Dissertation Kurzendorfer Tanja
Dissertation Kurzendorfer Tanja
Dissertation Kurzendorfer Tanja
zur
Tanja Kurzendorfer
aus
Amberg
Als Dissertation genehmigt von der
Technischen Fakultät der
Friedrich-Alexander-Universität Erlangen-Nürnberg
I would like to thank a number of people who have helped and supported me to write
this thesis, both regarding the content and the motivation.
First of all, I am very grateful to my supervisor Prof. Dr.-Ing. habil. Andreas Maier.
I appreciate his outstanding confidence in me, his support, and guidance over the
years, not only as a scientific mentor, his encouragement, and the freedom he allowed
me regarding the contents of my work. Prof. Maier reviewed my work carefully and
provided valuable feedback. His input helped to improve the quality of my work.
I deeply appreciate the support from Dr.-Ing. Alexander Brost and Dr.-Ing Christoph
Forman during the last years as my Siemens advisors. The discussions we had were
always encouraging and contributed to a continuous improvement of this work. They
spent a lot of time in reviewing papers and giving valuable feedback regarding the
methods.
Many thanks to my colleagues at the Pattern Recognition Lab, for the pleasant and
friendly atmosphere at the lab, for the ongoing knowledge sharing, and for the joyful
times outside the working hours. In particular, let me thank Katharina Breininger
and Peter Fischer for the great time and inspiring scientific discussions. Among the
students I have supervised, let me acknowledge Sabrina Reiml and Negar Mirshaz-
adeh for their excellent work that contributed to several publications.
Last but not least, I want to thank my family and friends for their patience and
support over the years. In particular, I would like to thank my boyfriend for the
steady support, the patience, and the encouragement over all the years.
Tanja Kurzendorfer
Contents
I Introduction 1
Chapter 1 Introduction 3
1.1 Heart Failure and Treatment Options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Scientific Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.3 Organization of the Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
i
Chapter 4 Left Ventricle Segmentation in 3-D LGE-MRI 59
4.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
4.2 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
4.3 Automatic Left Ventricle Segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
4.3.1 Registration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
4.3.2 Short Axis Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
4.3.3 Endocardial Contour Extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.3.4 Epicardial Contour Extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
4.3.5 Papillary Muscle Segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
4.3.6 Mesh Generation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
4.4 Semi-Automatic Left Ventricle Segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . 76
4.4.1 Smart Brush. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
4.4.2 Control Point Extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
4.4.3 Control Point Merging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
4.4.4 3-D Interpolation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
4.4.5 Surface Reconstruction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
4.5 Evaluation and Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
4.5.1 Data Description. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
4.5.2 Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
4.5.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
4.6 Discussion and Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
ii
IV Outlook and Summary 141
Appendix 149
Bibliography 169
iii
iv
PART I
Introduction
1
CHAPTER 1
Introduction
1.1 Heart Failure and Treatment Options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Scientific Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.3 Organization of the Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
According to a study from the world health organization in 2015, the major cause
of death worldwide is cardiovascular disease1 . Cardiovascular disease is a generic
term that includes several diseases, such as high blood pressure, coronary heart
disease, myocardial infarction (MI), stroke, congenital cardiovascular defects, and
heart failure (HF) [Go 14]. For diagnosis in clinical routine cardiac magnetic reso-
nance imaging (MRI) is used, as it can provide information on morphology, tissue
characterization, blood flow, or perfusion [Peti 11, Suin 14]. The technical progress in
the last years gave rise to advanced image analysis. One major interest is the viability
assessment of the myocardium and further analysis, which can be used for planning
and guidance of minimal invasive procedures. Late gadolinium enhanced (LGE)-MRI
(cf. Section 2.1.2) is clinically the gold standard for the assessment of myocardial vi-
ability [Rash 15]. The main research goal of this thesis is to develop segmentation
algorithms for robust segmentation of the left ventricle (LV), scar quantification, and
visualization, which can be used for planning and guiding during minimally invasive
procedures.
In this chapter, a short introduction on the medical background of congestive HF
and the development of myocardial scar is given. In addition, one treatment option
is described suitable for patients with advanced drug-refractory heart failure, systolic
dysfunction, and ventricular dyssynchrony. Furthermore, the achieved scientific con-
tributions regarding left ventricle segmentation, scar quantification, and visualization
are detailed. Finally, the outline of this thesis is presented.
3
4 Introduction
Rhythm
disorders Valve
Other
disorder of defects High blood
the heart
Heart pressure
problems Poor blood
Heart muscle
supply to
defects
lungs
Coronary
heart disease
Lung Lung disease, asthma
problems bronchitis, obstructed
airway
Figure 1.1: The common causes of heart failure are among others, heart problems,
high blood pressure, lung problems, lifestyle, infections, or other medical conditions.
The graphic is based on Ponikowski et al. [Poni 14].
tigue, exercise intolerance, diminished appetite, and depression [Shea 03]. The most
common cause of HF related to cardiac issues is a myocardial infarction [McMu 12].
According to Dickstein et al. [Dick 08] at least 51 % of patients suffering from HF have
an ischemic history. Myocardial infarction describes the ischemic cell death caused
by a disturbed perfusion of the heart cells [Tayl 05]. However, also other factors like
viral infections, personal lifestyle, heart problems, high blood pressure, lung prob-
lems, or other medical conditions may cause heart failure [Poni 14]. An illustration
of potential causes for HF is provided in Figure 1.1.
In general, heart failure is defined as the inability to generate sufficient cardiac
output to meet the metabolic needs of the organs and the tissue of the human body.
This can include systolic, diastolic, and high-output failure. If there are symptoms of
systematic or pulmonary fluid volume overload, this is considered as congestive heart
failure [Tayl 05].
The New York Heart Association (NYHA) classification is widely used to rate the
cardiac disease according to the symptoms in four different classes. Class I patients
have HF but no limitation of physical activity. Class II patients have mild symptoms,
which means they are comfortable at rest, but physical activity results in symptoms
like shortness of breath. Class III patients are comfortable at rest, but already show
symptoms at low level activity, for example walking for a short distance. Class IV
1.1 Heart Failure and Treatment Options 5
Table 1.1: NYHA functional classification for heart failure based on physical activity
and symptoms. This table is based on the ESC Guidelines [Dick 08].
Figure 1.2: (a) Normal ECG signal with a normal QRS complex. (b) Abnormal
ECG signal with a wide QRS complex.
patients already have symptoms at rest. See Table 1.1 for a detailed classification of
the NYHA classes.
Patients suffering from HF are commonly asked to estimate the distance they are
able to walk before getting out of breath. This is a measure of exercise capacity
and helps to classify the patients into the NYHA classes [Dick 08]. Patients at risk
for chronic HF should be screened for symptoms of f atigue, activity intolerance,
congestive symptoms, edema, and shortness of breath (FACES) [Tayl 05].
All patients that show symptoms for chronic HF should undergo further diag-
nostic evaluation of the LV function. Most important is the left ventricular ejection
fraction, which can be measured by electrocardiography or radionuclide ventriculog-
raphy [Tayl 05]. This information helps to distinguish systolic from diastolic dysfunc-
tion. If systolic dysfunction is verified, the duration of the QRS complex should be
investigated. If the duration is greater than 120 ms, it is often associated with a left
bundle branch block [Abra 03]. The delay of the electrical signal can be seen with an
electrocardiogram (ECG). In Figure 1.2 a normal ECG with a normal QRS complex
is compared to an abnormal ECG with a widened QRS complex.
The electrical signal is delayed by the left bundle branch, therefore the right ven-
tricle contracts earlier as the left ventricle. This leads to an asynchronous contraction
of the heart and reduces the pump efficiency of the left ventricle [Shea 03]. The left
ventricular ejection fraction can be less than 35 % [Shea 03, Abra 03, Blee 06]. Nor-
mally, the left ventricular ejection fraction is between 50 % and 70 % [Will 07]. In
general, about 30 % to 50 % of all patients that suffer from chronic heart failure also
have a preserved ejection fraction [Sant 14].
6 Introduction
Right
atrial lead CRT
generator
Coronary
sinus lead
Right
ventricular lead
Figure 1.3: Illustration of a CRT pacemaker implanted in the chest, with the LV
lead placed in the coronary sinus (green). The other leads are placed in the right
atrium (blue) and right ventricle (red).
at the same time. This bi-ventricular pacing should result in a narrower and more
normal QRS complex [Shea 03].
The problem, however, with cardiac resynchronization therapy is that about 30 %
to 40 % of the patients do not clinically respond to this therapy [Shet 14]. The ade-
quate patient selection is very important, as some patients do not benefit from CRT
implantation [Auri 11]. Furthermore, the number of non-responders may be related
to suboptimal lead positioning. Pacing the LV generally produces the best haemo-
dynamic response, but patients may not respond if the lead is placed on scar, as
it is not electrical conductive. Additionally, there is a significant variation in the
response when pacing different regions of the LV [Shet 12]. Finally, the therapy can
be optimized by adequate device programming and follow up of the patient after the
implantation [Brig 13].
Fluoroscopic imaging is used to guide the placement of the three leads, but this
modality provides neither functional nor anatomical information with respect to the
heart [Ma 12]. A typical fluoroscopic image of a CRT procedure is shown in Figure 1.4.
Due to the complexity of the procedure, the implantation of the LV lead through the
coronary vein may take long. This results in a considerable amount of fluoroscopic
exposure and a repeated use of contrast agent to outline the cardiac chambers and
great vessels [Bram 10]. In particular, the optimal placement of the LV lead in a
sub-branch of the coronary sinus is one of the most challenging aspects of the CRT
implantation. In addition, the presence of myocardial scar at the LV placement area
is an important determination of the response rate. If the lead is positioned in areas
with transmural scarring, the desired effect of CRT on clinical outcome and cardiac
performance may be reduced [Morg 09].
Integrating the knowledge of the LV’s anatomy, mechanical activation, and scar
tissue distribution to treatment planning and guidance is likely to improve the out-
come of the CRT implantation [Leyv 11, Beha 17, Moun 17]. Recent studies from
Shetty et al. [Shet 12] demonstrated that an overlay approach using MRI fused to
8 Introduction
X-ray is a powerful combination for planning and guiding a CRT procedure. The
mechanical activation of the heart wall can be estimated from MR, ultrasound, CT,
or C-arm CT [Jian 14, Moun 17, Kape 05, Po 11, Mull 14].
Cardiac MRI is the clinical gold standard for assessing the functional parameters
of the heart [McMu 12]. To visualize fibrosis and scarring LGE-MRI is used [Bako 13].
Besides technological improvements regarding image acquisition and the clear clinical
demand [Bilc 08], the challenge arises in providing automatic tools for image analysis.
Therefore, the main research goal of this thesis is to develop segmentation tools for
LGE-MRI sequences, scar quantification methods, and visualization options for 2-D
as well as 3-D LGE-MRI data sets, to support optimal procedure guidance.
An important goal of this thesis is to directly segment the left ventricle out of 2-D
and 3-D LGE-MRI to avoid errors from the cine MRI contour propagation. For the
2-D LGE-MRI the short axis (SA) scan is used for the segmentation of the contours.
Therefore, the LV is automatically detected and a morphological active contours with-
out edges approach is used to estimate the blood pool. Afterwards, the endocardial
and epicardial boundary can be estimated either using a filter based or learning based
method in combination with dynamic programming. For the endocardial approach,
an additional scar exclusion step is added. The contributions have been presented at
three international conferences [Kurz 17a, Kurz 17e, Kurz 18a].
A different segmentation pipeline is developed for 3-D LGE-MRI. As this scan
is a whole heart scan, first the left ventricle has to be initialized using a two-stage
registration step. In the next step, the short axis orientation is estimated using
principal component analysis. Having the rough outline of the left ventricle either
a learning based or a filter based approach can be applied for the segmentation of
the endocardial and epicardial border. For the endocardial boundary refinement, an
additional scar exclusion step is added. Furthermore, because of the high resolution
also the papillary muscles can be segmented from the 3-D LGE-MRI scan. The
methods were presented at international conferences [Kurz 15, Kurz 17b, Kurz 16a,
Kurz 17c] and in one international peer-reviewed journal [Kurz 17f].
In addition to the fully automatic left ventricle segmentation, an intuitive, in-
teractive, and generic segmentation tool is developed based on radial basis function
interpolation for the segmentation of the left ventricle. However, as it is a generic
segmentation tool, any region of interest can be segmented. First, a smart brush is
implemented that allows to perform fast 2-D segmentations of individual slices. In
the next step, from the segmented masks the contours are extracted and scattered
1.2 Scientific Contributions 9
data points are computed. From these points the 2-D normal vectors are estimated
based on the curvature of the contour. Afterwards, our new formulation of Hermite
radial basis functions is applied to reconstruct the 3-D surface. Effective 2-D interac-
tions using the smart brush and less number of equations accelerate the performance
of the 3-D interpolation. Thus this method allows a real-time and an intuitive seg-
mentation of 3-D objects and can handle any 3-D data set. The contributions have
been presented at one international conference [Mirs 17] and in one peer-reviewed
journal [Kurz 17d].
Scar Quantification
Having the segmented contours from the myocardium, the scar quantification can be
performed fully automatic. Therefore, the full-width-at-half-max (FWHM) and the
x-standard deviation (x-SD) algorithms are implemented in a fully automatic man-
ner, to be independent from any user interaction. Furthermore, a learning based scar
quantification is developed, to analyze the texture of the myocardium and not only
being dependent on the intensity distribution of the myocardium. These contribu-
tions were presented at two international conferences [Kurz 15, Kurz 18b] and in an
international scientific journal [Kurz 17f].
Scar Visualization
Having the segmented scar, different visualization methods can be applied. The
scar can be visualized as one 3-D surface mesh, however, with this visualization
it is not possible to delineate between endocardial or epicardial scar. Therefore,
a novel scar layer visualization is developed. The scar is divided into several lay-
ers to allow the user to delineate between endocardial, mid-cardial, and epicardial
scar. Furthermore, the transmurality of the scar can be evaluated by visual in-
spection. However, this method cannot just be applied to 3-D visualization, it can
also be used for the 2-D American Heart Association (AHA) bull’s eye plot (BEP)
(cf. Section 6.3.3). This novel visualization method has been presented at two con-
ferences [Reim 17a, Kurz 17g].
A serious of contributions to image guided therapy have been made, which are in
close correlation to this thesis.
One aim of the segmentation is to use the segmented objects for guidance dur-
ing the procedure. Therefore, the benefit of X-ray magnetic resonance fusion has
been investigated. Mainly three procedures are considered, guidance for congenital
heart disease catheterization procedures, guidance for bone biopsies, and guidance for
sclerotherapy [Kurz 14d, Kurz 14a, Gira 14, Kurz 14e, Hwan 14a, Hwan 14b, Kurz 14b,
Bao 14, Kurz 14c]. It has been shown, that using MRI overlay for interventional ra-
diology is sufficient. However, for more complex procedures, such as congenital heart
disease catheterization procedures 3-D meshes are beneficial, i. e., as also the ostia of
a vessel can be visualized.
10 Introduction
An additional research focus has been on the guidance of atrial fibrillation pro-
cedures. The goal of this procedure is to electrical isolate the pulmonary veins
from the left atrium [Bros 13]. One common catheter for the pulmonary vein iso-
lation is a cryo-balloon catheter. However, this catheter is hardly visible under
fluoroscopy. Therefore, special detection and tracking algorithms have been devel-
oped [Bour 12, Kurz 12, Kurz 13, Kurz 16b, Kowa 15].
In Chapter 2, the medical and technical background for this thesis is explained.
First, the basic principles of magnetic resonance imaging are described. Afterwards,
the difference between 2-D slice selective sequences and 3-D MRI volumes is outlined.
Finally, the basic principle of a LGE-MRI acquisition is introduced. In the second
section, the image processing basics used in this thesis are outlined, consisting of
the polar transform, a short overview of segmentation methods, and the active con-
tours without edges approach. In the last section, the pattern recognition pipeline
is depicted and the background which is needed in this work is explained. First, the
principles of a decision tree are detailed. Afterwards, the random forest classifier is
introduced and finally, the idea of cross-validation is depicted.
Late gadolinium enhanced MRI is the gold standard to visualize myocardial scar-
ring. The challenge of this MRI acquisition is the accurate segmentation of the my-
ocardium, as it is a prerequisite for most automatic scar quantification approaches.
In Chapter 3, segmentation approaches for 2-D LGE-MRI in SA orientation are in-
troduced. For the 2-D LGE-MRI, a filter based and a learning based approach is
described, respectively. For both approaches, the left ventricle is detected in the
short axis image stack. Afterwards, the endocardial boundary is estimated either
using a filter based detection or a learning based classification in combination with a
minimal cost path search in polar space. Having the endocardial contour, the epicar-
dial contour is refined. Afterwards, the data and the metrics used for the evaluation
are described. The proposed methods are evaluated on 100 clinical 3-D LGE-MRI
data sets.
In the last years, the imaging quality continuously improved and LGE-MRI was
extended to 3-D. This sequence allows for a more precise quantification of myocardial
scar.
However, little research has been done in the processing of these 3-D LGE-MRI
sequences. Therefore, in this work three novel methods, a filter based, a learning
based, and a semi-automatic approach, for the segmentation of the left ventricle in
1.3 Organization of the Thesis 11
3-D LGE-MRI are presented. For the filter based and learning based approach, the
left ventricle is detected using a two-stage registration approach. Second, the short
axis view of the left ventricle is estimated. Third, the endocardial contour is refined
by either using an edge detection or a machine learning approach in combination
with a minimal cost path search. Fourth, the epicardial contour is refined using the
information of the endocardium. Fifth, the papillary muscles (cf. Section 4.3.5) are
segmented. Finally, the contours are extracted as 3-D surface meshes.
In addition, a generic semi-automatic approach based on Hermite radial basis
function interpolation is used. First, individual slices are annotated in 2-D using the
smart brush. Second, the contours of the segmented masks are extracted, control
points are defined, and their normal vectors are computed. Third, the new formula-
tion of Hermite radial basis function is used to reconstruct the 3-D surface.
The proposed methods are evaluated on clinical 3-D LGE-MRI data sets from two
clinical sites.
Chapter 5 – Scar Segmentation in LGE-MRI
The LGE-MRI sequences are acquired to visualize the viability of the myocardium.
A prerequisite for the scar quantification is the accurate segmentation of the my-
ocardium. Having the segmented myocardium, different methods can be applied for
a fully automatic scar quantification. In this work, different fully automatically meth-
ods are implemented, such as the x-SD or the FWHM algorithm which are considered
as state-of-the-art scar segmentation approaches. Furthermore, a learning based ap-
proach using intensity and texture features is investigated. The scar quantification is
evaluated using 30 clinical 3-D LGE-MRI data sets.
Chapter 6 – Scar Visualization
Possible directions for future research and remaining challenges for clinical translation
of the presented methodology are presented in Chapter 7. The thesis concludes with
a summary of the medical and technical background and the scientific contributions
presented in this work.
12 Introduction
Chapter 1
Introduction
Chapter 2
Chapter 3 Chapter 4
Left Ventricle Segmen- Left Ventricle Segmen-
tation in 2-D LGE-MRI tation in 3-D LGE-MRI
Chapter 5
Scar Segmentation in LGE-MRI
Chapter 6
Scar Visualization
Chapter 7
Outlook
Chapter 8
Summary
In this chapter, the basics of MRI and the technical fundamentals of this thesis are
summarized. In the first section, the physical basics of MRI are described. Then, a
differentiation between 2-D slice selective sequences and 3-D volumes is given. Af-
terwards, LGE-MRI is explained, which is the gold standard to visualize myocardial
viability. In the second section, the image processing basics, such as the polar trans-
form or the morphological active contours without edges approach are explained. In
the last section, the pattern recognition principles used in this thesis for the LV seg-
mentation are introduced, including decision trees, the random forest classifier, and
the evaluation using cross-validation.
ω0 = γ||B0 || , (2.1)
13
14 Clinical and Technical Background
short TE long TE
short TR T1 weighting –
long TR PD weighting T2 weighting
Table 2.1: The influence of the repetition time (TR) and the echo time (TE) on the
weighting of the images and therefore, on the image contrast. The effects of TE and
TR result in different contrast for the MRI, T1 , T2 , or proton density (PD) weighted.
This table is based on Weishaupt et al. [Weis 08].
where γ is the gyromagnetic ratio. The larger the field strength B0 of the magnetic
field, the higher will be the precessional frequency. The spins can then be divided
into two components, the longitudinal magnetization, which is aligned with the mag-
netic field, and a transverse magnetization, which is perpendicular to the magnetic
field [Weis 08].
The alignment of the spins can be changed by a transverse magnetization in com-
bination with a radio frequency pulse (RFP), whose frequency correlates with the
Larmor frequency ω0 . After the excitation the spin realigns back to the magnetic
field B0 . This state is the so-called resting state, as this is the energetically best
position for the spins. Considering the two different components of a spin, the lon-
gitudinal magnetization will increase and the transverse magnetization will vanish.
The increase of longitudinal magnetization over time is called the T1 relaxation time.
T1 is tissue specific, i. e., a tissue with long T1 recovers slowly after a RFP. The van-
ishing of the transversal magnetization is associated with T2 relaxation. Tissues with
a short T2 will loose very fast the transverse magnetization. The contrast of the MRI
sequence is determined by the timing of the excitation pulses, the successive gradient
magnetic field, and the T1 and T2 relaxation processes [Dyma 04]. The parameters
that control the weighting are the echo time (TE) and the repetition time (TR). TE
is defined by the time delay after each RFP is measured, i. e., the loss in transverse
magnetization. Hence, TE defines the T2 weighting of the images. TR is the period
of time until the RFP is repeated, as several similar measurements need to be per-
formed to encode multiple lines of an image. The RFP will flip the excited spins
to a more transversal magnetization. The relaxation will recover the longitudinal
magnetization, which is determined by T1 . Furthermore, a third type of contrast can
be applied, which is called proton density (PD) weighting, where T1 and T2 weight-
ing are minimized. Therefore, only the variations due to proton density itself are
left [Weis 08]. In Table 2.1 the effects of TE and TR are summarized. In Figure 2.1
example images of the brain with the different contrast are shown.
In the following subsections, the difference between 2-D vs. 3-D sequences is out-
lined and LGE-MRI is introduced, which is used for the visualization of myocardial
scarring.
Figure 2.1: (a) T1 weighted image, where the contrast is based on a short TR and
a short TE. (b) T2 weighted image, where the contrast is based on a long TR and
a long TE. (c) PD weighted image, where the contrast is based on a long TR and
a short TE. The images are courtesy of the Siemens Healthcare GmbH, Erlangen,
Germany [Hend 03].
the z-direction. For more details about the additional phase encoding, please refer
to Weishaupt et al. [Weis 08]. The 3-D volume is computed by applying a 3-D recon-
struction after acquisition. In general, 3-D volume-selective sequences can achieve a
higher spatial isotropic resolution, however the scan time increases [Weis 08].
In the recent years, 3-D volume-selective acquisition have been applied for whole
heart imaging, where the whole heart is covered by a single acquisition. However,
a long acquisition time was the drawback of the sequences. Therefore, special sam-
pling patterns, such as the Cartesian spiral phyllotaxis sampling pattern has been
introduced for a faster data acquisition [Picc 11]. Furthermore, free breathing acqui-
sitions have been developed by applying respiratory gating [Picc 12, Grim 13, Form 13,
Form 14, Form 15, Wetz 17]. These technologies have the potential to bring 3-D car-
diac imaging into clinical routine.
Late gadolinium enhanced MRI is the clinical gold standard for non-invasive assess-
ment of myocardial viability [Kim 00, Kell 12, Shin 14, Rash 15]. The contrast of this
sequence is based on gadolinium (Gd), which belongs to the lanthanide series of
the chemical elements, which are often referred as the rare earth elements [Weis 08].
Gadolinium in its pure form is a paramagnetic, toxic metal and therefore, an un-
favorable element in the human body. Hence, gadolinium is never injected in the
pure form in the body but bounded to a carrier molecule. The purpose of the car-
rier molecule, called chelating agent is to modify the distribution of the gadolinium
within the human body, to overcome the toxicity but maintaining the contrast prop-
erties. Different vendors use different chelating molecules, to bind the paramagnetic
gadolinium Gd3+ [Rogo 16]. The contrast agent is injected intravenously and will be
eliminated later trough the kidneys.
LGE-MRI uses the difference in contrast agent accumulation between viable and
damaged tissue, which leads to various enhancements within the myocardium. The
increased concentration of the contrast agent in the scar tissue is based on a larger
extra-vascular, extra-cellular volume, and slower washout. LGE-MRI images are
typically acquired 10 min to 20 min after a gadolinium based contrast agent injec-
tion [Kim 00, Kell 12]. The enhancement is based on T1 -shortening and the different
distribution of the contrast agent. GRE-based inversion recovery (IR) or phase sensi-
tive inversion recovery using ECG gating are widely used sequences for the acquisition
of LGE-MRI. The sequence has normally a 180◦ IR pre-pulse to flip the signals of all
tissues to the negative axis. The inversion time (TI) is different for each patient, as
it is dependent on the patient weight, contrast dose, renal function, and time-point
after the contrast agent injection. After the inversion pulse, the tissue recovers back
to the resting state. The signal should be measured exactly when the healthy my-
ocardium tissue crosses the zero line. Hence, the resulting image depicts the normal
myocardium as dark tissue and scar tissue is visualized bright. Figure 2.2 depicts the
acquisition protocol of an LGE-MRI sequence. It can be seen that the data is ECG
gated to mid-diastole of every other cardiac cycle. Furthermore, the tissue specific
TI is visualized, which is normally around 300 ms [Dyma 04].
2.1 Magnetic Resonance Imaging 17
ECG
Acquisition
window
Trigger
Viable myocardium
T [ms]
TI
300 ms
Mz
Recently, the use of IR with single shot steady state free precession, known as
true-FISP, has become more popular. This sequence has a fast multi-slice coverage
and provides good results for patients who suffer from arrhythmias or have difficulty
holding their breath [Kell 12]. In clinical routine, 2-D multi-slice LGE-MRI has proven
to provide diagnostic accuracy. The acquisition of a single slice is acquired during the
period of one breath hold. The slice selective breath hold scan has to be repeated up to
14 times, till the entire LV is covered in the desired orientation [Shin 14]. Commonly,
the SA view is used for the LV. In Figure 2.3 (a) the SA orientation and in (b) the
long axis (LA) orientation of the SA scan is visualized for a 2-D LGE-MRI sequence,
where the myocardial scarring is highlighted through an orange arrow.
Recently, LGE-MRI was extended to 3-D to continuously cover the whole heart
with a high resolution in a single acquisition. This scan guarantees an accurate
quantification of the myocardium to the extent of myocardial scarring. However,
the acquisition time can be increased, because it is a free breathing scan which is
navigator gated. But no resting periods between the scans have to be included for
the patient’s comfort [Shin 14]. To increase the acquisition time of the scan the
acquisition is improved using compressed sensing [Form 13]. Figure 2.3 depicts a 3-D
LGE-MRI sequence, where (c) corresponds to the SA orientation and (d) to the LA
view, where the myocardial scar is highlighted through orange arrows.
18 Clinical and Technical Background
Figure 2.3: Comparison of a 2-D and 3-D LGE-MRI of the left ventricle, showing
two individual patients. (a) 2-D LGE-MRI SA slice of the LV, with a myocardial
infarction (orange arrow). (b) LA view of the 2-D data, with a bad inter-slice reso-
lution because of the large slice thickness and distance between the slices. The 2-D
LGE-MRI sequence has the following parameters: pixel spacing (1.77 × 1.77) mm2 ,
slice thickness 8 mm, spacing between the slices 10 mm, number of slices 10. (c) 3-D
LGE-MRI reorientated in a pseudo SA view. (d) 3-D LGE-MRI in a pseudo LA
view. With the 3-D LGE-MRI an accurate quantification to the extent of myocardial
scarring is possible, the orange arrows are pointing to the MI. The 3-D LGE-MRI
sequence has the following parameters: voxel size 1.30 mm3 isotropic, number of slices
120.
After describing the fundamentals for the image acquisition, the basic image process-
ing methods used in this thesis are described.
First, the polar transform is detailed, as the polar space is widely used for left
ventricle segmentation [Joll 11, Cord 11, Huan 11, Drei 13, Hu 13, Hu 14]. Second, a
general overview of segmentation methods is given. Third, a morphological active
contours without edges approach is described, which is applied for the 2-D left ven-
tricle segmentation for the rough estimation of the blood pool.
2.2 Image Processing 19
where r ∈ R+ ◦ ◦
0 is the radius and ρ ∈ [0 , 360 ] is the angle in polar space. The
conversion is straight forward for continuous functions. However, if the function is
defined as a discrete and equally spaced grid, such as an image, an additional bilinear
interpolation is needed for the conversion [Park 07]. In this thesis, the origin of the
polar space image corresponds to the center of the endocardium, see Figure 2.4.
For the final result of the left ventricle segmentation, the polar coordinates are
transformed back to Cartesian coordinates
x r · cos(ρ)
pc = = . (2.3)
y r · sin(ρ)
79
0 79
x
38
0◦ 90◦ 180◦ 270◦ 360◦
ρ
Figure 2.4: (a) Image in Cartesian coordinates, where the center of the blood pool
is marked with an orange dot. (b) Polar image of (a), where the origin of the image
corresponds to the center of the blood pool in Cartesian space.
Model based methods have been successfully used in the last years for image
segmentation [Heim 09, Mont 01]. Deformable models are physically motivated, as
they deform a closed parametric curve under the influence of internal and external
energy functions [Pham 00]. Kass et al. [Kass 88] first introduced freely deformable
models as snakes using explicit deformable contours. The evolution of the curve is
driven by two forces, an external energy that adapts the model to the image and an
internal energy that stabilizes the shape based on smoothness criteria. The contour
S called active contour or snake is evolved by minimizing the energy functional F .
The energy functional F is given by
Z 1 Z 1 Z 1
0 2 00 2
F (S) = µ |S (p)| dp + ν |S (p)| dp − λ1 |∇I(S(p))|2 dp , (2.4)
0 0 0
where µ, ν, and λ1 ∈ N are positive constants that control the strength of each
term and ∇I is the gradient image of the input image I. The first two terms of
Equation (2.4) control the smoothness of the contour, i. e., the internal energy and
2.2 Image Processing 21
the third term attracts the contour towards the boundary of the image, i. e., the
external energy [Chan 01].
However, parametric deformable models have two main limitations. First, the
initialization of the model is very crucial. Second, topological adaption such as split-
ting or merging of model parts is very difficult, as a new parametrization must be
constructed [Ange 05]. Osher and Sethian [Oshe 88] introduced the concept of geo-
metric deformable models. This formulation can be used as an implicit formulation
of the deformable contour, as it allows for cusps, corners, and automatic topological
changes [Chan 01]. Hence, geometric deformable models were introduced to address
these limitations [Case 93], using the mean curvature motion equation [Oshe 88]
∂u ∇u
= |∇u|div , (2.5)
∂t |∇u|
where u : Rd → R is the defined level set at time t. Therefore, the energy functional
of Equation (2.4) is given by
∂u ∇u
= |∇u|div e(I) , (2.6)
∂t |∇u|
where the image gradient ∇I is replaced by a general edge function e(I) [Crem 07].
This new formulation is also known as geodesic active contours, as the energy can be
interpreted as length of the contour [Case 97].
Nonetheless, the stopping of the curve is still depending on the image gradient. To
overcome this issue Chan and Vese [Chan 01] introduced a new formulation of active
contours, which is not dependent on the edge function, the so-called active contours
without edges (ACWE). The external energy function is based on the Mumford-Shah
segmentation [Mumf 89], where the image is modeled as a piecewise-smooth function.
The energy functional F takes the content of the interior Ω and exterior Ω regions of
the curve S into account. The functional of the curve S is dependent on the image
I ∈ RN ×M and defined by
Z Z
F (c1 , c2 , S) = µ·L(S)+ν ·A(S)+λ1 ||I(p)−c1 ||2 dp+λ2 ||I(p)−c2 ||2 dp, (2.7)
Ω Ω
of MACWE is that the stopping of the curve evolution is not dependent on the edge
function, instead it is region-based, which means that it uses image statistics both
inside and outside of the contour. The difference to ACWE [Chan 01] is that this
approach is based on morphological discretization instead of partial differential equa-
tions, as they are computationally expensive and may have stability issues, depending
on the initialization [Marq 14].
The goal is to minimize the energy functional F of Equation (2.7),
For a fixed contour S, c1 and c2 are the mean intensity values of the area inside Ω, and
outside Ω of the curve S. The third and the fourth term of the energy functional F
of Equation (2.7) minimize the intensity difference of the values inside and outside of
the contour S. The first and second term of the energy functional F of Equation (2.7)
are added for regularization purposes [Chan 01].
The Euler-Lagrange equation for the implicit version of the functional F in Equa-
tion (2.7) using a level set formulation is [Chan 01]
∂u ∇u 2 2
= |∇u| µ · div − ν − λ1 (I − c1 ) + λ2 (I − c2 ) , (2.9)
∂t |∇u|
where u : Rd → R is the defined level set at time t . This equation specifies how the
curve S should evolve to minimize the functional F in a steepest descent manner. It
consists of the balloon term, the smoothing term, and the image attachment term.
The balloon force is given by the equation
∂u
= ν|∇u| . (2.10)
∂t
Depending on the sign of ν, the balloon term is equivalent to the classical morpho-
logical operators, erosion and dilation
i
D(u (p))
if ν > 0
ui+1 (p) = E(ui (p)) if ν < 0 , (2.11)
i
u (p) if ν = 0
where D is the dilation, E is the erosion, and i is the iteration step. The smoothing
term is given by the equation
∂u ∇u
= |∇u|µ div . (2.12)
∂t |∇u|
∂u
= |∇u| λ1 (I − c1 )2 + λ2 (I − c2 )2 .
(2.13)
∂t
2.3 Pattern Recognition 23
Figure 2.5: (a) Initialization of the MACWE algorithm. (b) Result of the MACWE
evolution.
The attraction force term depends on the mean intensities inside c1 and outside c2 of
the curve S [Marq 14]
1
if λ1 (I − c1 )2 < λ2 (I − c2 )2
ui+1 (p) = 0 if λ1 (I − c1 )2 > λ2 (I − c2 )2 . (2.14)
i
u (p) otherwise
See Figure 2.5 for an exemplary result for the MACWE in a 2-D LGE-MRI sequence,
where the non-negative parameters are set as follows: µ = 1, ν = 1, λ1 = 1, and
λ2 = 2.
Classification Phase
I Pre- Feature f ci
Recording Classification
Processing Extraction
• other classifiers: random forest, support vector machine, hidden Markov mod-
els, neural networks, etc.
Important to know is, that every classifier is only as good as the training samples
used. In addition, the more training samples available the better the classifier will
generalize. In the next two sections, the focus in on decisions trees and the random
forest classifier, as they are used in this thesis for boundary classification.
In 1984, Breiman et al. [Brei 83] first introduced a tree model for classification
and regression, which is also known as classification and regression trees (CART).
From then on, decision trees became very popular and were widely used for different
machine learning problems. Reasons for their success are that they are fast and
scalable even for large data sets and they can be formulated in a probabilistic fashion.
Breiman [Brei 01] further developed the ensemble of decision trees and introduced in
2001 the injection of randomness in the learning phase of each individual tree. Since
then, the random forest has been successfully applied to many different machine
learning applications.
In this section, the basic machine learning methods used in this thesis are sum-
marized. First, the principles of decision trees are described and their application
for classification is introduced. Second, the concept of the random forest classifier is
explained. Afterwards, the idea of cross-validation is described and the optimization
of the hyper-parameters of the random forest classifier is explained.
2.3 Pattern Recognition 25
→ Root Node
→ Split Nodes
→ Leaf Nodes
Figure 2.7: A decision tree consists of a set of nodes, which are connected through
edges. The top node is referred to as root. The internal nodes or split nodes are
visualized as circles. The leaf or also called terminal nodes are depicted as rectangles.
In contrast to graphs, in a decision tree there are no loops.
where 0 and 1 represent the decision for the left and right node and θ is a set of
parameters for the split function. The parameters of θ can be further decomposed
into
θ = (φ, ψ, τ ) , (2.16)
0 0
where φ : Rd → Rd , with Rd << Rd is the feature selector function, which selects
some features of choice out of the vector f . The parameter ψ ∈ Rd defines the
geometric primitive used for separating the data, i. e., axis-aligned hyper-plane or
an oblique hyper-plane. The parameter τ saves the threshold used in the binary
26 Clinical and Technical Background
x2 ψ x2 ψ
x1 x1
(a) Axis-aligned hyper-plane (b) General oriented hyper-plane
Figure 2.8: Examples of splitting functions, which are also referred to as week
learners. Mostly, linear splitting functions are used, followed by a thresholding op-
eration. (a) An axis aligned hyper-plane as splitting function. (b) General oriented
hyper-plane as splitting function. For visualization purposes the feature vector is in
2-D space, f = (x1 , x2 )T ∈ R2 . Images based on Criminisi et al. [Crim 12].
test [Crim 13]. All the parameters will be optimized at each split node [Crim 12]. In
Figure 2.8 different examples of splitting functions are visualized. Commonly, linear
functions coupled with a threshold operation are applied
During the training of the decision tree, the optimal parameters for θ i for each of
the split nodes needs to be computed. Considering classification, the aim is to reduce
the class uncertainty. The most popular objective functions are the information
gain and a variation of it called Gini impurity. The information gain needs to be
maximized as it measures the difference between the class uncertainty before and
after the splitting,
right
θ ?i = argmax Gi (Di , Dleft
i , Di , θi ) , (2.18)
θi
where G : D → R is the information gain for one node. The information gain after
the split is computed by
X |Di |
G = H(D) − H(D) , (2.19)
|D|
i∈{1,2}
where c ∈ C indicates the class label, C ∈ {c1 , ..., cN } is the set of all classes, and p
the empirical distribution of the training point of c within the test data D.
2.3 Pattern Recognition 27
1.0
0.8
Class percentage
0.6
x2
0.4
0.2
0.0
x1 Class distribution
0.8 0.8
Class percentage
Class percentage
0.6 0.6
x2
0.4 0.4
0.2 0.2
0.0 0.0
x1 Class distribution Class distribution
Figure 2.9: Information gain after splitting. (a) Data distribution before splitting,
where the different colors indicate different classes. (b) Class distribution before
splitting, with a uniform distribution, as for every class the same number of points is
given. (c) Data distribution after vertical splitting. (d) Class distribution after the
vertical split for the left part, only the red and blue class are left and the distribution
is uniform. (e) Class distribution after the vertical split for the right part, only the
green class is left. Graphic based on Criminisi et al. [Crim 12].
Figure 2.9 (a) depicts three different classes, where the color indicates the class
label. The distribution over the classes is uniform, because each class has the same
number of points, which can be seen in Figure 2.9 (b). In Figure 2.9 (c) the data
is split vertically, which produces two sets of data, Dleft and Dright . Each data set
after the split has higher information and therefore, a lower entropy. Therefore, the
information gain is maximized after the first split, as one class is already separated.
The calculation of the information gain happens during the training phase of the
forest. The decision functions at each node are optimized individually to iteratively
split the training data until a stopping criteria is reached [Crim 12].
There are mainly three common stopping criteria: i) maximal tree depth D ∈ N is
reached, ii) minimum population per leaf is reached, or iii) minimum variation of the
information gain G. If one of these criteria is fulfilled, the last tree node is reached,
the iterative splitting of the training data D stops and the node will turn into a leaf
node.
The function of the leaf node is to model the posterior distribution of the given
subset of D of the training data. Hence, each leaf node corresponds to a part of
28 Clinical and Technical Background
x2
x1
Figure 2.10: (a) Partition of feature space into three classes. (b) Resulting decision
tree, where each leaf node represents one class.
the feature space, as depicted in Figure 2.10. The probabilistic leaf model can be
formulated as
p(c|f ) , (2.22)
where c ∈ C is indexing the class label and p ∈ [0, 1] is the probability [Crim 12].
To further speed up the decision process and avoid overfitting, leaf pruning can
be applied [Crim 13]. The approach after Gelfand et al. [Gelf 89] proposes to split
the training data set into two disjunct sets and alternating the role of the both sets.
First, the tree is grown using the first half and pruned using the error rate of the
second half. Afterwards, the tree is regrown from the pruned tree using the previous
test half and pruned using the previous training half. These steps are repeated until
the tree size does not change anymore.
To conclude this section, decision trees can approximate new information if the
training data is large enough. However, a single decision tree is prone to overfitting.
Therefore, in the next section the idea of the random forest classifier is introduced to
overcome overfitting.
x2
x1
Figure 2.11: (a) Randomly sampled feature space. (b) Three individual trees,
trained with a randomized subset of the feature space.
However, a pre-requisite for a well-performing forest model is that all trees are
randomly different from each other. The randomness is achieved during the train-
ing phase, where mainly two different approaches are used. i) The training data is
sampled randomly, which was defined as bagging by Breiman [Brei 01], as it is a com-
bination of bootstrap and aggregating. Given the training data set D a subset Dt is
extracted, where all the elements have been randomly sampled using a uniform distri-
bution. The splitting and randomization of the training data yields a good training
efficiency. In Figure 2.11 an example of the randomized splitting of the feature space
D is given, where each of the three trees is trained on a different subset Dt of the
training data. ii) A randomized node optimization is applied [Geur 06], where each
tree can be trained on the whole training data D. For the ith node of a tree, the set
of all possible parameters θ can be defined as P. However, during the training only
a small random subset P i ⊂ P of the possible parameters is used. This leads to an
individual adopted optimization function for each node
Underfitting Overfitting
Prediction error
Prediction error
Figure 2.12: The number of the trees T and the tree depth D are the two most im-
portant parameters regarding the optimization of the random forest. (a) Correlation
between the number of trees T and the convergence of the prediction error. (b) Cor-
relation between the depth of the tree D and the problem of over- or underfitting the
data.
2.3.3 Cross-Validation
For the training of a classifier, parameters have to be learned. If the testing would
be performed with the training data set, the model would have a very good predic-
tion score, e. g., for the nearest neighbor classifier the labels would just be repeated.
However, this is a methodological mistake and is called overfitting. To avoid such
situations, the data set D has to be split into a training set Dtrain and a test data set
Dtest , where D = Dtrain ∪ Dtest and Dtrain ∩ Dtest = ∅.
Furthermore, overfitting can be present when the hyper-parameters of a classifier
are optimized on the test set Dtest , therefore, the evaluation metric does no longer
report the general performance of the classifier. Hence, another data set has to be ex-
tracted from the training set Dtrain , the so-called validation set Dvalidation , where Dtrain
2.3 Pattern Recognition 31
= D0train ∪ Dvalidation . The training is then performed on the training set D0train , after-
wards, the evaluation of the hyper-parameters is done on the validation set Dvalidation .
After satisfying results are achieved with the training, the final evaluation can be
done using the test data set Dtest .
However, if there is just a limited amount of data available, the splitting of the
data drastically reduces the number of samples available in each set and the results
can depend on the choice of the data sets. There is a need for a large training,
validation, and test set at the time to achieve a generic classifier which is not sensitive
to the training and validation data and to prove the generalization, i. e., the subject
independence by a large test data set. To solve this problem, a common approach
is to use cross-validation (CV), or also often called k-fold CV, where the data set is
split into k disjoint subsets of equal size. The model is trained on k − 1 subsets of the
data set D. Afterwards, the trained classifier is evaluated using the remaining part
of the data set D where the classification error is estimated. The steps are repeated
k times, until every subset was used for testing. In Figure 2.13 (b) an illustration
for a k-fold CV is given, where k = 3. Using a k-fold CV the classification error
can be estimated for the whole data set D, and more precise statements about the
performance of the classifier can be given [Cawl 10]. In the extreme case, where k
corresponds to the size of the data set D, the cross-validation is called leave-one-out
CV. In Algorithm 2.1, a pseudo algorithm for the cross-validation is presented.
Cross-validation can be combined with a grid search for the optimal hyper-parameter
selection of the classifier, as the hyper-parameters, such as the tree depth D and the
number of trees T for the random forest, are not directly learned within the training
phase of the classifier. A grid search can be used to estimate the performance of all
32 Clinical and Technical Background
different parameter combinations for one classifier. Therefore, a range of the possible
parameters has to be specified. With a nested k-fold CV, the hyper-parameters and
the evaluation of the classifier can be achieved at the same time [Cawl 10]. For the
nested cross-validation, two loops of cross-validation are required. The outer loop
splits the data D into a training data set Dtrain and a testing data set Dtest . The
inner loop splits the training data Dtrain of the outer loop into a disjoint training
D0train and validation data set Dvalidation . Within the inner loop the hyper-parameter
of the model are optimized by a cross-validation combined with a grid search. After
the optimal parameters have been found, the outer loop is trained with the hyper-
parameters and evaluated using the test set Dtest . In Algorithm 2.2, the individual
steps of the nested cross-validation are detailed. If the model is stable and also the
features used for training the model are well selected, the hyper-parameters θ i for
each of the k-folds should be similar. However, in real world scenarios, the parame-
ters are often different for each split. Then for the final model, i. e., the mean value
of each parameter can be used for training.
PART II
33
CHAPTER 3
This chapter presents methods for the left ventricle segmentation in 2-D LGE-MRI.
In particular, filter based and learning based methods for the LV segmentation are
introduced. This chapter is organized as follows: The motivation is depicted in
Section 3.1. In Section 3.2 relevant literature is reviewed. The proposed segmentation
framework is introduced and described in Section 3.3. The evaluation and results
using 2-D LGE-MRI are shown in Section 3.4. In Section 3.5 the results are discussed
and a conclusions is drawn. Parts of this chapter have previously been published in
three conference publications [Kurz 17a, Kurz 17e, Kurz 18a].
35
36 Left Ventricle Segmentation in 2-D LGE-MRI
3.1 Motivation
The clinical gold standard to visualize myocardial scarring is 2-D LGE-MRI [Born 95,
Shin 14, Rash 15]. In clinical routine, the left ventricle’s myocardium in 2-D LGE-MRI
is commonly segmented manually. However, manual delineation of the myocardial
boundary is complex, time consuming, and prone to intra- and inter-observer variabil-
ity. The main challenge with the segmentation of LGE-MRI is the non-homogeneous
contrast distribution within the myocardium due to the different contrast agent accu-
mulation in the healthy and damaged myocytes. Therefore, the challenge lies in the
accurate contour delineation of the endocardial and epicardial boundary of the left
ventricle. Hence, most 2-D LGE-MRI segmentation methods rely on the registration
of the cine MRI scans to the LGE-MRI. In Figure 3.1 the workflow of the cine MRI
registration to the LGE-MRI is depicted.
The registration and fitting from the cine MRI to the LGE-MRI has several prob-
lems. The global position of the heart may change during the acquisitions due to
patient movement, as there is a large time gap between the acquisition of the two
sequences. The cardiac phases from the cine MRI may not precisely match to the
LGE-MRI. Inter-slice shifts can arise from multiple breath-holds. Even though these
shifts may appear minor, they can lead to significant errors in the myocardial scar
quantification.
Dikici et al. [Diki 04] use a non-rigid variational approach to register cine MRI
with LGE-MRI. Afterwards, active contours are used to optimize a parametric affine
transform of the propagated parameters. Cifolo et al. [Ciof 08] propose to deform a
geometrical template to fit to the myocardial contour for each MRI slice. In addition,
the LV was divided into four quadrants and those likely to contain large areas of
myocardial scarring were treated differently. The contours from the cine MRI were
then aligned to the geometrical template of the LGE-MRI. Wei et al. [Wei 11] use a
model based approach for the LV segmentation that comprises several steps. Inter-
slice shifts in cine MRI images are corrected and the LGE-MRI is registered to the
cine MRI. The cine MRI contours are further deformed by features in the LGE short
or long axis images. Tao et al. [Tao 14] propose a four step algorithm to segment
the LV from LGE-MRI. The LGE-MRI is aligned to the cine MRI. The endo- and
epicardial borders from cine MRI are fitted to the LGE-MRI, optimized in 3-D and
refined based on the LGE-MRI.
Alabà et al. [Alba 14] propose a modified graph cut approach to segment the LV
using 2-D LGE-MRI. Therefore, six rules are defined: (i) the blood pool appears
bright, (ii) the blood pool has a circular shape, (iii) the blood pool is continuous in
the longitudinal direction, (iv) the myocardium can include dark and bright voxels,
(v) the myocardial thickness changes smoothly, and (vi) the 3-D global shape of the
myocardium is smooth. The method from Alabà et al. does not use information
from cine MRI. The direct segmentation using LGE-MRI avoids segmentation errors
caused by the use of the 2-D cine MRI and a misalignment between the sequences. In
addition, it allows for a faster segmentation approach. Hence, the direct segmentation
of the LV solely based on LGE-MRI is desirable. The main difference from Alabà
et al. to the proposed method is that Alabà et al. use a graph cut approach, in
combination with shape and interslice constraints, which require prior information
derived from the six rules. In the proposed algorithm, a morphological active contours
approach is used for the rough estimation of the blood pool. For the boundary
estimation either a filter based or learning based approach is used in combination
with a minimal cost path search in polar space.
The segmentation of the LV can be divided into four major steps. First, the left ven-
tricle is detected using a combination of circular Hough transforms, Otsu’s thresh-
olding, and a circularity measure (Section 3.3.1). In the second step, a region of
interest is identified using a morphological active contours without edges (MACWE)
approach (Section 3.3.2). In the third step, the endocardial boundary is estimated
(Section 3.3.3). In this thesis two approaches are described, a filter based segmenta-
tion and a learning based segmentation [Kurz 17a, Kurz 17e, Kurz 18a]. In the last
step, the epicardial contour is obtained by considering the endocardial boundary
(Section 3.3.4). Figure 3.2 provides an overview of the segmentation pipeline.
38 Left Ventricle Segmentation in 2-D LGE-MRI
Figure 3.2: Overview of the LV segmentation pipeline in 2-D LGE-MRI. First, the
LV is detected using circular Hough transforms and a circularity measure. In the
next step, the blood pool is segmented by applying a MACWE approach. In the
third step, the endocardial border is refined in polar space using a dynamic program-
ming approach. In the final step, the epicardial contour is extracted considering the
endocardial contour and the edge information.
The LV is detected in the mid-slice of the 2-D LGE-MRI stack. First, the Canny
edge detector is used to extract the edges from the image [Cann 86], as depicted in
Figure 3.3 (a). In the next step, circular Hough transforms are applied [Duda 72].
The radii of the circular Hough transforms are in range of 17 mm to 35 mm, which
is defined according to the anatomical information in literature [Lang 06] and with
a step size of 2 mm due to performance. Figure 3.3 (b) illustrates the five most
prominent circles which are detected using the circular Hough transform. The most
prominent candidate is selected as potential LV blood pool candidate. To verify this
position, an additional roundness measure of different objects is estimated. Therefore,
Otsu’s thresholding is applied to the whole slice, to convert the image into a binary
mask [Otsu 79], see Figure 3.3 (c). Afterwards, binary erosion is applied and objects
that are smaller than a predefined threshold θo ∈ R, where θo = 25, are removed.
The resulting image is visualized in Figure 3.3 (d). The threshold for θo is defined
heuristically. Fromqthe remaining objects the eccentricity, i. e., the roundness R ∈ R
2
is estimated R = 1 − ab 2 , where a is the semi-major axis and b is the semi-minor
axis of the object. If the object is circular, R = 0. In Figure 3.3 (e) the roundness
measure for the remaining three objects is calculated, where prior to the calculation
the convex hull of the objects is estimated. If the center points c1 and c2 of the
roundest object and the result of the circular Hough transform are within θc ∈ R,
where θc = ||c1 − c2 ||2 , the LV has been accurately detected. And the center point c
of the blood pool is defined as c = 12 (c1 + c2 ). Otherwise, the user is asked to verify
the center of the LV by clicking into the center. The result of the LV detection is
shown in Figure 3.4 (a).
3.3 Segmentation Pipeline 39
Figure 3.3: LV detection pipeline. (a) Canny edge image. (b) Result of the circular
Hough transform, showing the five most dominant circles. (c) Binary image after
Otsu’s threshold was applied. (d) Erosion and small object removal. (e) Results
of the roundness measure for the three remaining objects, where the convex hull is
applied for the computation.
After the center of the LV is detected in the center slice of the LGE-MRI stack,
this information is used for the boundary estimation of the endocardium. The mid-
slice is a good slice to start with the segmentation, as the result is used to propa-
gate in basal and apical direction. Having the rough location of the LV, the center
point is used as initialization for a morphological active contours approach with-
out edges (MACWE) [Marq 14]. A detailed introduction of the MACWE method is
provided in Section 2.2.3.
The functional for the contour S ∈ R2 is dependent on the image slice I ∈ RN ×M
and defined by
Figure 3.4: Mid-slice of the LGE-MRI stack. a) Detected center of the LV using
circular Hough transforms and circularity constraints. (b) Result of the morphological
active contours without edges approach (MACWE).
After the blood pool approximation of the LV, the contour S is used for the refinement
of the endocardial border. As the rough outline of the LV is known, the image I is
cropped around the region of interest to perform further image processing steps on
the cropped image.
The polar image is calculated, where the origin of the polar image corresponds to
the center of the blood pool. The maximum radius is selected to cover all potential
myocardium boundaries. The estimated contour S from the MACWE approach is
also converted to polar coordinates and used to refine the endocardial boundary.
Figure 3.5 (a) depicts an example of the polar image and the converted contours.
In the polar image the edge information is extracted by applying the Canny edge
detection [Cann 86]. To extract minor edges, a Gaussian smoothing with a standard
deviation of σ = 1.5 is used, see Figure 3.5 (b) for an example. In addition, the
mean intensity µbp ∈ R and the standard deviation σbp ∈ R of the blood pool are
estimated using the intensity values inside the contour obtained from the MACWE
approach. In the next step, a scar threshold θst ∈ R is defined as θst = µbp + σbp . All
pixel intensities that are greater than θst and outside of the blood pool are defined as
potential myocardial scar and labeled with 1, the non-scar pixels are labeled with 0,
see Figure 3.5 (c). Afterwards, all pixels with increasing radius after the potential scar
candidates are labeled with 1, resulting in a scar map, as visualized in Figure 3.5 (d).
The scar map is combined with the edge image, which derives the cost array K ∈ Rr×ρ
for the minimal cost path search, as shown in Figure 3.5 (e). Six equally distributed
points are selected from the converted blood pool contours S and the MCP search is
3.3 Segmentation Pipeline 41
0
r [pixel]
38
0◦ 90◦ 180◦ 270◦ 360◦
ρ
(a) Polar image with contours
0
r [pixel]
38
0◦ 90◦ 180◦ 270◦ 360◦
ρ
(b) Canny edge image
0
r [pixel]
38
0◦ 90◦ 180◦ 270◦ 360◦
ρ
(c) Scar probability
0
r [pixel]
38
0◦ 90◦ 180◦ 270◦ 360◦
ρ
(d) Scar map
0
r [pixel]
38
0◦ 90◦ 180◦ 270◦ 360◦
ρ
(e) Cost array
0
r [pixel]
38
0◦ 90◦ 180◦ 270◦ 360◦
ρ
(f ) Final minimal cost path (MCP) result
Figure 3.5: Final steps of the endocardial boundary segmentation. (a) Polar image
with the converted contour S from the blood pool segmentation in red. (b) Edge
image using the Canny edge detection. (c) Scar classification, all pixel that are greater
than θst and outside of the blood pool are labeled with 1. (d) Scar map, where all
pixels with increasing radius after a scar candidate are labeled with 1. (e) The cost
array is derived from the edge image combined with the scar map. (f) Final result of
the minimal cost path search.
42 Left Ventricle Segmentation in 2-D LGE-MRI
initialized to find the optimal contour [Dijk 59]. The MCP finds the distance weighted
minimal cost path through the cost array. The cost path is calculated as the sum
of the costs at each point of the path, where edge pixels have 0 costs and non edge
pixels 1. The cost c ∈ R of one step from point pi to pi+1 is calculated as follows:
d d
c (pi , pi+1 ) = · K (pi ) + · K (pi+1 ) , (3.2)
2 2
where d ∈ R is the distance between the two points pi and pi+1 . The diagonal
versus axial moves are of different length and therefore, the path costs are weighted
accordingly. The final contour C ∈ RN ×D of the MCP is illustrated in Figure 3.5 (f).
After the refinement, the contour C is converted back to Cartesian coordinates.
As the result might be frayed and papillary muscles close to the endocardial border
may be included, the convex hull is calculated for the estimated contour points C.
The final contour C is derived from the smallest convex polygon, as depicted in the
third box of Figure 3.2. This assumption is based on the fact that the left ventricle’s
cavity is convex in the SA view.
After the first contour is refined, the information of the location of the LV is used
for the initialization of the MACWE for the succeeding slices in basal and apical
direction. The refinement steps are repeated for every slice. Furthermore, the radius
and area of the previous curve are considered for the refinement of the lower and
upper slices.
The refinement is stopped in apical direction if either the last slice is reached, or
the radius of the endocardial contour r falls below a threshold θr ∈ R.
For the refinement in basal direction the iteration stops, if either the last slice is
reached, the difference of the areas of the previous contour C i−1 and the newly found
contour C i is greater than a threshold θbmax ∈ R, or the left ventricular outflow tract
is detected. For the detection of the left ventricular outflow tract, the root mean
square error of the intensities of the current slice to the previous slice is calculated.
If the error is greater than θLVOT ∈ R, the left ventricular outflow tract is reached
and the refinement stops.
As for the filter based approach, the outline of the MACWE approach is used to
extract potential endocardial boundary candidates using circular ray casting. There-
fore, the image is converted to polar coordinates, as for the filter based approach.
Potential boundary candidates are then selected for N ∈ N equidistant points along
the contour S, as depicted in Figure 3.6 (a) in Cartesian coordinates. Each potential
boundary candidate is then classified using a trained random forest (RF) classifier.
The result of the classification is illustrated in Figure 3.6 (b) and (c) as a cost map,
where green corresponds to 0 costs and red corresponds to 1.
The performance of any classifier is limited by the discriminative power of the
features used for training. Steerable features are used [Zhen 08], as they are compu-
tationally efficient and can capture the orientation and scale. In total, 16 features are
extracted for each boundary candidate, based on local intensity and gradient, which
result in a feature vector f ∈ R16 that is used for training and detection. For a given
boundary candidate p = (x, y)T with the intensity I ∈ R and the gradient g ∈ R2 ,
3.3 Segmentation Pipeline 43
1.0
0.8
0.6
0.4
0.2
0.0
Figure 3.6: Learning based boundary classification using a random forest classifier.
(a) Potential boundary candidates extracted using circular ray casting. (b) Endocar-
dial boundary detection result obtained from the trained RF classifier. (c) Boundary
cost map in Cartesian coordinates.
√ √
where 5I(p) =pg = (gp T
x , gy ) , the following features are extracted:
p I, I, 3 I, I 2 ,
I 3 , log I, ||g||, ||g||, 3 ||g||, ||g||2 , ||g||3 , log ||g||, gx , gy , gx 2 + gy 2 , and 52 I(p).
Note, that all the features are extracted in polar space, which is the steerable space.
The center position in Cartesian space, i. e., the origin in polar space, has no influence
on the classification result.
The training of the RF is based on ground truth annotations from which positive
as well as negative samples are extracted. For the training pathologic as well as
healthy subjects are used to generate a broad range for the training data base. After
the training, the classifier predicts the endocardial boundary probability pendo (f ) ∈
[0, 1]. The endocardial boundary probability can be interpreted as costs c, where
c = 1 − pendo . If the boundary probability is very high, the costs are close to 0.
To improve the detection of the boundaries in areas of scarred myocardium, an
additional scar exclusion step is added. Given the mean intensity of the blood pool µbp
and the standard deviation σbp , the scar threshold θst is defined as θst = µbp + σbp .
All the pixels above this threshold and outside of the blood pool are defined as
potential scar candidates, see Figure 3.7 (c). The scar map is generated from the scar
candidates, where all pixels with increasing radius from potential scar candidates
are labeled with 1, as depicted in Figure 3.7 (d). In the next step, the boundary
probability is compared to the scar map, at locations where the scar map is labeled
with 1, the boundary map is also labeled with 1. Hence, the boundary potentials are
impaired, which results in the endocardial cost array K ∈ Rr×ρ , see Figure 3.7 (e).
In the next step, the final segmentation result of the endocardial contour has to
be obtained from the endocardial cost map K. Therefore, a dynamic programming
approach is used in polar space, to compute the optimal endocardial contour from
left to right of the polar image [Qian 15]. The minimal cost path (MCP) search is
used [Dijk 59], which finds the distance weighted minimum path through the cost ar-
ray, as for the filter based approach. The result of the MCP is shown in Figure 3.7 (f).
After the optimal path is found, the contour is transferred back to Cartesian
coordinates, see Figure 3.8 (a). The convex hull is calculated from the contour, as
44 Left Ventricle Segmentation in 2-D LGE-MRI
0
r [pixel]
38
0◦ 90◦ 180◦ 270◦ 360◦
ρ
(a) Polar image
0 1
r [pixel]
0.5
38 0
0◦ 90◦ 180◦ 270◦ 360◦
ρ
(b) Boundary costs
0 scar
r [pixel]
38 healthy
0◦ 90◦ 180◦ 270◦ 360◦
ρ
(c) Scar candidates
0 1
r [pixel]
38 0
0◦ 90◦ 180◦ 270◦ 360◦
ρ
(d) Scar map
0 1
r [pixel]
0.5
38 0
0◦ 90◦ 180◦ 270◦ 360◦
ρ
(e) Boundary costs combined with scar map
0
r [pixel]
38
0◦ 90◦ 180◦ 270◦ 360◦
ρ
(f ) MCP result
Figure 3.7: Final steps of the learning based endocardial boundary segmentation.
(a) Mid-slice image after polar transformation. (b) Boundary cost map obtained from
the trained RF classifier in polar coordinates. (c) Potential scar candidates which
have an intensity value greater than θst and are not within the blood pool. (d) Scar
map, where all scar candidates with increasing radius are labeled with 1. (e) Final
endocardial boundary cost map, resulting from the combined boundary detection
result and the scar map. (f) Final result of the minimal cost path (MCP) search in
polar coordinates.
3.3 Segmentation Pipeline 45
Figure 3.8: Final steps of the endocardial boundary segmentation. (a) Result of
the minimal cost path search (MCP) in Cartesian coordinates. (b) Convex hull of
the result shown in (a).
For the epicardial contour extraction the previously found endocardial contour points
are used. The contour extraction is performed in polar space. As the epicardium
has to be greater than the endocardium, the radius is enlarged by θepi ∈ R, see Fig-
ure 3.9 (a) for an example, where the red contour corresponds to the final endocardial
contour and the yellow contour to the enlarged endocardial contour. The previously
calculated edge image from the endocardial refinement is used. The edges within
the enlarged endocardial segment are erased. Figure 3.9 (b) depicts the edge image
which is compared to the enlarged endocardial mask in Figure 3.9 (c) and results
in an image where only possible epicardial edges remain and all edges within the
endocardial mask are set to 0, see Figure 3.9 (d). Having the enlarged endocardial
contour, the closest edge with increased radius, in a certain margin is searched for, as
46 Left Ventricle Segmentation in 2-D LGE-MRI
0
r [pixel]
38
0◦ 90◦ 180◦ 270◦ 360◦
ρ
(a) Enlarged endocardial contour
0
r [pixel]
38
0◦ 90◦ 180◦ 270◦ 360◦
ρ
(b) Edge image
0
r [pixel]
38
0◦ 90◦ 180◦ 270◦ 360◦
ρ
(c) Endocardial mask
0
r [pixel]
38
0◦ 90◦ 180◦ 270◦ 360◦
ρ
(d) Epicardial edges
0
r [pixel]
38
0◦ 90◦ 180◦ 270◦ 360◦
ρ
(e) Detected edges for epicardial contour in yellow
Figure 3.9: Final steps of the filter based epicardial boundary segmentation. (a) Po-
lar image with the smoothed endocardial contour in red and the enlarged endocardial
contour in yellow. (b) Edge imaged derived from the Canny edge detector. (c) En-
larged endocardial binary mask. (d) The resulting epicardial edge image is compared
to the edge image in (b) and the enlarged endocardial mask in (c), which results in an
edge image showing only the remaining possible epicardial edges. (e) With increased
radius the next edge to the enlarged endocardial contour with increasing radius is
searched for, resulting in the yellow contour.
3.4 Evaluation and Results 47
Figure 3.10: Final steps of the epicardial boundary segmentation. (a) The final
result of the boundary estimation for the endocardium in red and the epicardium in
yellow. (b) 3-D model of the endocardial and epicardial contour in red and yellow,
respectively.
illustrated in Figure 3.9 (e). Afterwards, the newly found edge points are converted
back to Cartesian coordinates. As the result is frayed, the convex hull is estimated
to smooth the contour and remove outliers.
After the first slice is refined, the steps are repeated till the base and apex are
reached, i. e., the segmentation for the epicardial contour stops at the same slice as
the endocardial contour refinement is stopped.
For the learning based approach, also the endocardial contour is used as a starting
point. From the enlarged contour, potential boundary candidates are extracted using
circular ray casting. Again a RF classifier is trained for the epicardial boundary de-
tection using the same 16 features as for the endocardial border estimation, resulting
in an epicardial boundary probability pepi (f ). The result of the epicardial boundary
detection is used as cost array for the MCP search. The MCP is applied in polar
coordinates for the same reasons as mentioned before. The MCP finds the distance
weighted minimal path from the left to the right end of the polar image. The result
is then transferred back to Cartesian coordinates and the convex hull is taken. The
result is depicted in Figure 3.10 (a).
For all the segmentation approaches, the contours of the myocardium are extracted
as 3-D surface models using the marching cubes algorithm [Lore 87]. For the marching
cubes algorithm a standard scikit-image Python implementation is used. The output
is a list of vertices and faces, which are automatically saved as a standard surface
mesh format. Figure 3.10 (b) shows an example of a 3-D surface mesh, where the
endocardium is visualized in red and the epicardium in yellow.
section, the evaluation methods are presented. In the last section, the results are
shown for the two segmentation approaches and the two approaches are compared to
each other.
3.4.2 Evaluation
Given the gold standard annotations, the segmentation algorithms are evaluated us-
ing different measures: the volumetric Dice coefficient (DC) and the mean surface
distance (MSD). The DC is a quantitative score to estimate the segmentation qual-
ity, as it measures the proportion of true positives in the segmentation result. The
DC ranges from 0 to 1, with 1 corresponding to a perfect overlap. The metric is
defined as
2|A ∩ B|
DC(A, B) = , (3.3)
|A| + |B|
where A is the segmentation result, which corresponds to a set of pixel and B is
the gold standard annotation from one of the clinical experts. The MSD calculates
the mean surface distance between the surface voxel of the binary mask A and their
nearest surface voxel of the gold standard mask B, averaged over all contour points.
Furthermore, the inter-observability between the two clinical experts is evaluated.
3.4.3 Results
In the first paragraph, the results of the filter based segmentation are presented.
In the second paragraph, the results of the learning based segmentation are shown.
Finally, both results are compared against each other.
The filter based segmentation is evaluated on 100 clinical data sets, which are detailed
in Section 3.4.1. The automatic segmentation of the endocardium results in an overlap
with the gold standard segmentation of 0.80 ± 0.11 considering the Dice coefficient.
The best segmentation result achieves a DC of 0.95 and the worst a DC of 0.47. For
the epicard, a mean DC of 0.81 ± 0.09 is achieved. The best segmentation result of
3.4 Evaluation and Results 49
Endo Epi
Description Mean ± Std Min Max Mean ± Std Min Max
Mean 0.80 ± 0.11 0.47 0.95 0.81 ± 0.09 0.35 0.95
Observer 1 0.79 ± 0.11 0.47 0.95 0.80 ± 0.10 0.35 0.93
Observer 2 0.81 ± 0.11 0.47 0.94 0.82 ± 0.08 0.56 0.95
Inter-Observer 0.95 ± 0.06 0.61 0.99 0.96 ± 0.05 0.68 0.99
Table 3.2: Quantitative segmentation results of the filter based 2-D LGE-MRI
segmentation using the DC. The results are shown separately for the endocardial
(Endo) and epicardial (Epi) contour. Furthermore, the mean DC for both observers,
as well as the mean DC for each of the observer is presented. In addition, the inter-
observer variability between the two observers is investigated.
the epicardium yields a DC of 0.95 and the worst a DC of 0.35. As two gold standard
annotations are available, also the inter-observer variability is evaluated. The inter-
observer variability between the two observers results in a DC of 0.95 ± 0.06 for the
endocardium and 0.96 ± 0.05 for the epicardium. The best overlap for the inter-
observer variability achieves a DC of 0.99 for the endocard and epicard, respectively.
The worst inter-observer variability DC for the endocard results in 0.61 and in 0.68
for the epicard. In Table 3.2 the results for the Dice coefficient are summarized and
the two different observers are distinguished, as well as the inter-observer variability
is presented. In Figure 3.11 the mean Dice coefficient for every data set is visualized
separately for the endocard and epicard. The endocard is sorted in increasing order
according to the Dice coefficient. The epicard is sorted according to the endocard.
It can be seen that the DC of the endocard and epicard in general correlates well,
which results in a person correlation coefficient of 0.82. The major difference between
the endocard and epicard is in the apex, as there the boundary delineation for the
epicard is sometimes hard to see.
Furthermore, the mean surface distance is evaluated. For the endocardium a mean
distance of 3.71 mm ± 2.57 mm is accomplished, with a minimum distance of 0.44 mm
and a maximum distance of 11.79 mm. For the epicard a mean distance of 4.33 mm
± 2.65 mm is achieved with a minimum distance of 0.5 mm and a maximum distance
of 16.34 mm. As previously, the MSD for the inter-observer variability is evaluated,
which results in a mean distance of 0.89 mm ± 1.15 mm between the two observers
for the endocardium and 0.92 mm ± 1.14 mm for the epicardium. The minimum
distance for the inter-observer variability results in 0.00 mm for the endocardium and
epicardium, respectively. The maximum inter-observer variability distance for the
endocardium is 6.01 mm and for the epicardium 5.55 mm. In Table 3.3 the results
for the mean surface distance are shown. In addition, the two different observers are
distinguished, and the inter-observer variability is summarized.
For the evaluation of the learning based approach, a nested cross-validation is used,
as described in Section 2.3.3. The nested cross-validation is needed for the random
forest classifier as hyper-parameters need to be optimized. As these parameters are
50 Left Ventricle Segmentation in 2-D LGE-MRI
0.9
0.8
Dice Coefficient
0.7
0.6
0.5
0.4
Endocard
0.3
1 10 20 30 40 50 60 70 80 90 100
Sequence Number (sorted)
0.9
0.8
Dice Coefficient
0.7
0.6
0.5
0.4
Epicard
0.3
1 10 20 30 40 50 60 70 80 90 100
Sequence Number (sorted according to (a))
Figure 3.11: Individual Dice coefficients for the 2-D filter based segmentation for
each of the 100 data sets for the endocard and epicard, respectively. The DC for the
endocard is sorted in increasing order. The DC for the epicard is sorted according to
the endocard.
3.4 Evaluation and Results 51
Table 3.3: Quantitative segmentation results of the filter based 2-D LGE-MRI
segmentation using the MSD in mm. The results are shown separately for the endo-
cardial (Endo) and epicardial (Epi) contour. Furthermore, the mean MSD for both
observers, as well as the mean MSD for each of the observer is presented. In addition,
the inter-observer variability between the two observers is investigated.
not learned during the normal training phase of a classifier. For this purpose, a grid-
search is used, as it exhaustively compares all parameter combinations. For the 100
clinical data sets a 5-fold nested cross-validation is used. Hence, 20 data sets are
used for the testing of the classifier and the rest is used to train and validate the
classifier. For the grid-search the following parameter sets are evaluated: number of
trees T ∈ {30, 40, 50, 60, 70, 80}, maximal tree depth D ∈ {5, 10, 15, 20, 25, 30}, mini-
mum number of samples for the split node S ∈ {15, 20, 25, 30, 35, 40}, and minimum
number of samples required at the leave node F ∈ {2, 3, 4, 5, 6, 7, 8, 9}. The inner
loop of the nested cross-validation is also set to a 5-fold cross-validation. The op-
timal hyper-parameters for the random forest for the endocardium are summarized
in Table 3.4 (a) for each of the five folds. The optimal hyper-parameters for the
epicardium are summarized in Table 3.4 (b).
In Table 3.5 and Table 3.6, the results for the learning based approach are pre-
sented using the Dice coefficient and the mean surface distance, respectively. The
computed metrics are presented for the endocardium and epicardium, as well as for
the two different observers.
In Figure 3.12 the mean Dice coefficient for every data set is visualized separately
for the endocard and epicard. The endocard is sorted in increasing order according
to the Dice coefficient. The epicard is sorted according to the endocard. It can be
seen that the DC of the endocard and epicard in general correlates well, which results
in a person correlation coefficient of 0.69.
Furthermore, the feature importance for the endocard and epicard is evaluated,
see Figure 3.13 (a) and (b), respectively. It can be seen that the importance of the
features for the endocard and the epicard corresponds. In general, the gradient fea-
tures are more important compared to the intensity features. This can be attributed
to the non-homogeneous intensity distribution of the LGE-MRI in case of myocardial
scarring.
As the hyper-parameters for each fold are different, as shown in Table 3.4 (a) and
Table 3.4 (b), an additional evaluation is performed to see the impact of the different
parameters on the final segmentation result. Hence, the evaluation is repeated using
a regular 5-fold cross-validation for each of the 5 optimal hyper-parameter sets Hi
for the endocard and epicard, respectively. The results are shown in Table 3.7 for
the endocard and in Table 3.8 for the epicard. Furthermore, the mean Dice coeffi-
52 Left Ventricle Segmentation in 2-D LGE-MRI
0.9
0.8
Dice Coefficient
0.7
0.6
0.5
0.4
Endocard
0.3
1 10 20 30 40 50 60 70 80 90 100
Sequence Number (sorted)
0.9
0.8
Dice Coefficient
0.7
0.6
0.5
0.4
Epicard
0.3
1 10 20 30 40 50 60 70 80 90 100
Sequence Number (sorted according to (a))
Figure 3.12: Individual Dice coefficient for 2-D learning based segmentation for
each of the 100 data sets for the endocard and epicard, respectively. The DC for the
endocard is sorted in increasing order. The DC for the epicard is sorted according to
the endocard.
3.4 Evaluation and Results 53
Description T D S F
1
Hendo 80 15 3 20
2
Hendo 60 15 4 40
3
Hendo 70 15 8 40
4
Hendo 80 15 4 30
5
Hendo 50 15 6 15
(a) Optimized hyper-parameters for endocard
Description T D S F
1
Hepi 70 15 5 35
2
Hepi 80 20 5 40
3
Hepi 70 15 9 30
4
Hepi 70 20 9 35
5
Hepi 50 15 3 40
(b) Optimized hyper-parameters for epicard
Table 3.4: Optimized hyper-parameters for 2-D random forest, for each individual
fold. For the grid-search the following parameter sets are optimized: number of trees
T , maximal tree depth D, minimum number of samples for the split node S, and
minimum number of samples required at the leave node F .
Endo Epi
Description Mean ± Std Min Max Mean ± Std Min Max
Mean 0.82 ± 0.08 0.60 0.96 0.81 ± 0.08 0.36 0.95
Observer 1 0.81 ± 0.08 0.60 0.96 0.80 ± 0.09 0.36 0.94
Observer 2 0.82 ± 0.08 0.61 0.96 0.82 ± 0.07 0.58 0.95
Table 3.5: Quantitative segmentation results of the learning based 2-D LGE-MRI
segmentation using the DC. The results are shown separately for the endocardial
(Endo) and epicardial (Epi) contour. Furthermore, the mean DC for both observers,
as well as the mean DC for each of the observer is presented.
cient for each set of the hyper-parameters Hi for the endocardium and epicardium is
illustrated in Figure 3.14. It can be seen that changing the hyper-parameters has a
small influence on the final segmentation result, as the classification result is not the
final result for the boundary delineation. For the endocardium, an additional scar
exclusion step is added and afterwards the boundary is delineated using a minimal
cost path search for the endocardium and epicardium.
In Figure 3.15, the qualitative results of the segmentation approaches are presented.
Therefore, an average segmentation result is chosen. The first row depicts the raw
data from base to apex. The second row shows the gold standard annotation of
the first clinical expert i. e., observer 1, where the endocardial contour is orange and
the epicardial contour green. The third row illustrates the result of the filter based
54 Left Ventricle Segmentation in 2-D LGE-MRI
Table 3.6: Quantitative segmentation results of the learning based 2-D LGE-MRI
segmentation using the MSD in mm. The results are shown separately for the endo-
cardial (Endo) and epicardial (Epi) contour. Furthermore, the mean MSD for both
observers, as well as the mean MSD for each of the observer is presented.
0.20 0.20
Importance
Importance
0.10 0.10
0.00 0.00
q ||g||
q ||g||
log(I)
log(I)
log(||g||)
log(||g||)
I
I
I
I
gx
2
gx
2
3
2
3
gy
gy
q 5 I(p)
q 5 I(p)
||g||
||g||
||g||
||g||
√
√
||g||
q||g||
||g||
q||g||
y
y
2 + g2
2 + g2
I
I
3
3
2
2
3
3
gx
gx
Feature Feature
Figure 3.13: Feature importance of the random forest classifier for the 2-D
LGE-MRI classification. (a) Feature importance for the random forest classifier
trained for the endocardial boundary detection. (b) Feature importance for the
random forest classifier trained for the epicardial boundary detection. It can be seen,
that the feature importance correlates for both classifiers.
segmentation algorithm, where the endocardium is red and the epicardium yellow.
The last row visualizes the results of the learning based approach using the same
colors as for the filter based approach. It can be seen that the presented results
match well with the gold standard annotation. The biggest difference occur in the
apex, especially for the epicard.
In Figure 3.16, a comparison of the DC for the filter based vs. the learning based
segmentation of the endocardium and epicardium is given, respectively. It can be
seen that the learning based segmentation outperforms the filter based segmentation
and is more robust with respect to outliers, especially for the endocardium. However,
the final difference considering the DC is not that big, as the post-processing steps
are similar for both methods.
Furthermore, the correlation between the filter based and the learning based seg-
mentation was investigated using a scatter plot. The results of the scatter plot are
depicted in Figure 3.17 (a) for the endocard and in Figure 3.17 (b) for the epicard.
The Pearson correlation coefficient between the endocard of the two methods results
3.4 Evaluation and Results 55
all D1 D2 D3 D4 D5
1
Hendo 0.82±0.08 0.82±0.08 0.82±0.09 0.80±0.10 0.81±0.07 0.80±0.06
2
Hendo 0.82±0.08 0.87±0.07 0.82±0.09 0.79±0.10 0.81±0.07 0.81±0.06
3
Hendo 0.81±0.09 0.85±0.07 0.82±0.09 0.76±0.11 0.81±0.07 0.80±0.06
4
Hendo 0.81±0.09 0.87±0.07 0.82±0.08 0.78±0.11 0.80±0.07 0.80±0.07
5
Hendo 0.81±0.09 0.86±0.08 0.82±0.09 0.78±0.10 0.81±0.07 0.80±0.07
Table 3.7: Influence of hyper-parameters for 2-D endocardial random forest, for
each individual fold, using the DC as measure.
all D1 D2 D3 D4 D5
1
Hepi 0.81 ± 0.08 0.84 ± 0.07 0.82 ± 0.07 0.75 ± 0.11 0.82 ± 0.05 0.80 ± 0.07
2
Hepi 0.81 ± 0.08 0.84 ± 0.07 0.82 ± 0.06 0.75 ± 0.11 0.82 ± 0.06 0.80 ± 0.07
3
Hepi 0.81 ± 0.08 0.84 ± 0.07 0.82 ± 0.06 0.74 ± 0.11 0.82 ± 0.05 0.81 ± 0.07
4
Hepi 0.80 ± 0.09 0.84 ± 0.07 0.83 ± 0.06 0.74 ± 0.12 0.82 ± 0.06 0.80 ± 0.08
5
Hepi 0.80 ± 0.09 0.84 ± 0.07 0.82 ± 0.07 0.74 ± 0.12 0.82 ± 0.05 0.80 ± 0.08
Table 3.8: Influence of hyper-parameters for 2-D epicardial random forest, for each
individual fold, using the DC as measure.
1.0 1.0
Dice Coefficient
Dice Coefficient
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0.0 0.0
1
Opti Hendo 2 3 4 5 Opti 1 2 3 4 5
Hendo Hendo Hendo Hendo Hepi Hepi Hepi Hepi Hepi
Figure 3.14: (a) Comparison of the influence of the different sets of hyper-
i
parameters Hendo on the Dice coefficients for the endocardium, where the blue line
represents the mean Dice coefficient. (b) Comparison of the influence of the different
i
sets of hyper-parameters Hepi on the Dice coefficients for the epicardium, where the
blue line represents the mean Dice coefficient.
in 0.75 and in 0.80 for the epicard. Thus, there is a good correlation between the two
segmentation methods.
The proposed approaches are implemented on an Intel i7-4810MQ with 2.80 GHz
CPU equipped with 16 GB of RAM in Python 2.7. The runtime for both approaches
is very fast with around 10 seconds for one data set. However, the evaluation of the
random forest classifier using a nested cross-validation takes about 2 days. Solely
the training without optimization of the hyper-parameters takes about 1 minute,
depending on the forest size T , the tree depth D, and the parallelization of the
random forest.
56 Left Ventricle Segmentation in 2-D LGE-MRI
Figure 3.15: Comparison of the segmentation result for one data set, which rep-
resents an average segmentation result. From top to bottom: native slices without
any contours, gold standard annotation from first clinical expert i. e., observer 1, seg-
mentation result of the filter based segmentation, and segmentation result for the
learning based method.
The presented methods solely use 2-D LGE-MRI data for the segmentation of the
left ventricle, compared to most work reported in literature, which make use of cine
MRI and propagate the contours to the LGE-MRI [Diki 04, Wei 11, Ciof 08, Tao 14].
Alabà et al. [Alba 14] compute directly the contours from LGE-MRI. The presented
results are in the same range of the reported errors in literature. However, a direct
comparison to the method is not possible as the data sets differ.
3.5 Discussion and Conclusion 57
1.0 1.0
Dice Coefficient
Dice Coefficient
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0.0 0.0
Filter Based Learning Based Filter Based Learning Based
Figure 3.16: (a) Comparison of the Dice coefficient between the filter based and the
learning based segmentation for the endocardium, where the blue line represents the
mean Dice coefficient. (b) Comparison of the Dice coefficient between the filter based
and the learning based segmentation for the epicardium, where the blue line repre-
sents the mean Dice coefficient. It can be seen that the learning based segmentation
performs better with less outliers, both for the endocardium and epicardium.
1.0 1.0
DC Learning Based
0.8 0.8
DC Learning Based
0.6 0.6
0.4 0.4
0.2 0.2
0.0 0.0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
DC Filter Based DC Filter Based
Figure 3.17: Scatter plot of the Dice coefficient for the endocard and epicard of
the filter based segmentation compared to the learning based segmentation. (a) The
Pearson correlation coefficient for the endocard results in 0.75. (b) The Pearson
correlation coefficient for the epicard results in 0.80.
The proposed method achieves a DC of around 0.82 for the endocardium and
epicardium considering the learning based approach. The biggest differences occur in
the basal region, as the delineation of the left ventricular out flow tract is not always
clear. The poor performance of the MSD is mainly due to the larger error in the
apex and the left ventricular out flow tract. However, the results in the mid-cavity
are convincing, which is shown in Figure 3.15. It is expected that incorporating a
model will directly improve the segmentation result, especially in the apex and base.
In the course of this work, it has been shown that rather simple methods can be
used for the boundary detection of the endocardium and epicardium. In combination
58 Left Ventricle Segmentation in 2-D LGE-MRI
with a minimal cost path search, accurate and consistent results can be achieved.
The clear benefit of the method is the independence of registration to cine MRI and
the speed.
CHAPTER 4
In the previous chapter, methods for left ventricle segmentation in 2-D LGE-MRI are
presented. In this chapter, the segmentation of the left ventricle in 3-D LGE-MRI
is detailed. This chapter is organized as follows: The motivation is illustrated in
Section 4.1. Related literature is reviewed in Section 4.2. The automatic pipeline
for the LV segmentation using 3-D LGE-MRI is described in Section 4.3. A semi-
automatic approach for the left ventricle segmentation based on Hermite radial ba-
sis function (HRBF) is outlined in Section 4.4. The experimental results for 3-D
LGE-MRI are shown in Section 4.5. The results are discussed in Section 4.6 and a
conclusion is given. Parts of this chapter have previously appeared in five conference
publications [Kurz 15, Kurz 16a, Kurz 17b, Kurz 17c, Mirs 17] and two peer-reviewed
journal publications [Kurz 17f, Kurz 17d].
59
60 Left Ventricle Segmentation in 3-D LGE-MRI
4.1 Motivation
The viability assessment of the myocardium after a myocardial infarction is very
important for diagnosis, therapy planning, success prediction of the therapy, and pa-
tient prognosis. Especially, the location and quantification of a patient’s scar burden
promises to increase the success rate of different therapies, such as ablation of ven-
tricular tachycardia or CRT [Kari 16, Diki 04, Moun 17]. LGE-MRI is the clinical
gold standard for non-invasive assessment of myocardial viability [Shin 14, Rash 15,
Kell 12].
Recently, LGE-MRI was extended to 3-D to continuously cover the whole heart
with a high resolution of up to (0.7 × 0.7 × 1.5) mm3 in a single acquisition. This
scan promises an accurate quantification of the myocardium to the extent of my-
ocardial infarction [Shin 14], see Figure 4.1 for an example. Besides the technological
improvements regarding image acquisition and the clear clinical demand [Bilc 08],
the challenge arises in providing automatic tools for fast image segmentation and
analysis.
The main issue with processing LGE-MRI data is the non-homogeneous intensity
distribution within the myocardium, resulting from the different accumulations of
contrast agent in the damaged tissue. Moreover, the 3-D whole heart acquisition
does not allow a direct use of any ring or circular shape prior, as commonly used for
2-D SA images. Furthermore, the left ventricle is covered now in more than 80 slices
compared to around 10 slices for the 2-D LGE-MRI acquisitions. To overcome these
4.2 Related Work 61
Figure 4.1: (a) 3-D LGE-MRI reorientated in a pseudo SA view. (b) 3-D LGE-MRI
in a pseudo LA view. With the 3-D LGE-MRI an accurate quantification to the extent
of myocardial scarring is possible. The 3-D LGE-MRI sequence has the following
parameters: (1.30 × 1.30) mm2 pixel spacing, 1.30 mm slice thickness, 120 slices.
issues, two novel approaches for fully automatic left ventricle segmentation using 3-D
LGE-MRI data are proposed. The two different approaches, filter and learning based,
are detailed in the following sections.
is refined in polar space using a cost array and applying a minimal cost path search.
The cost array can either be derived filter based or learning based and always con-
siders potential scar candidates [Kurz 17f, Kurz 17c] Fourth, the epicardium is com-
puted, starting from the endocardium and either considering the edge information
for the filter based approach or the cost array for the learning based approach. Prior
knowledge, such as shape and inter-slice smoothness constraints, is used during the
refinement step. Finally, the endocardial and epicardial contours are exported as 3-D
surface meshes using the marching cubes algorithm [Lore 87]. Figure 4.2 provides an
overview of the segmentation pipeline.
4.3.1 Registration
For the initialization of the LV within the 3-D whole heart scan, a two-stage registra-
tion to an atlas volume A ∈ RN ×N ×N is performed [Unbe 15]. The atlas volume A has
a corresponding labeled mask L ∈ RN ×N ×N showing the left ventricle’s endocardium,
which was segmented manually. The registration can be defined as a one-to-one map-
ping between coordinates in one space and those in another space such that points
in the two spaces that correspond to the same anatomical structure are mapped to
each other [Maur 93]. In the first step, a multi resolution affine registration is per-
formed to roughly align the atlas volume A to the input volume V ∈ RN ×N ×N , see
Figure 4.3 (b) for an example registration result. The transformation T ∈ RN ×N can
be decomposed into scaling s ∈ R, rotation R ∈ RN ×N , and translation t ∈ RD ,
However, to match the complex deformations and anatomical changes between indi-
vidual patients, more variations need to be allowed. Hence, a non-rigid multi resolu-
tion registration is applied. The checkerboard result of the non-rigid registration is
depicted in Figure 4.3 (c). Nevertheless, an affine transform has to be applied prior
to the non-rigid registration for an initialization, as non-rigid transformations most
often cannot handle large scaling, translation, or rotation.
4.3 Automatic Left Ventricle Segmentation 63
Figure 4.3: (a) Checkerboard image before the registration, showing the thorax in a
sagittal viewing plane. (b) Checkerboard image showing the registration result after
the affine registration. (c) Final result after the non-rigid registration.
N
1 X
Σ= (v i − v̄) (v i − v̄)T , (4.2)
N i=1
1
http://elastix.isi.uu.nl/
64 Left Ventricle Segmentation in 3-D LGE-MRI
Figure 4.4: (a) 3-D LGE-MRI of the three orthogonal imaging planes, sagittal, coro-
nal, and axial. (b) Approximated cavity of the LV through the registration to the atlas
data A. The principal components are visualized, with the first principal component
in blue and the second and third principal components in black. (c) Cropped pseudo
short axis (SA) view after PCA. (d) SA view with the contour of the transformed
mask M in red.
After a pseudo SA view of the LV is estimated, the algorithm starts with the slice that
corresponds to the center of mass of the transformed mask M . The contour points
v i ∈ R3 of the transformed atlas mask M are extracted in the SA orientation using
the marching cubes algorithm [Lore 87]. In Figure 4.4 (d) the slice that corresponds
to the center of mass is depicted, with the contour of the transformed mask M in
red.
In the next step, the polar image is calculated, where the origin of the polar
image corresponds to the mean of the found contours v̄, i. e., the center of mass.
For the particular slice l ∈ {1, 2, ..., N }, Figure 4.5 (a) is the analogue image of
Figure 4.4 (d) in polar space. The maximum radius r ∈ R+ 0 is selected to cover all
potential myocardium boundaries and the angle ρ ∈ [0 , 360◦ ]. The contour points
◦
v i from the transformed mask are converted to polar coordinates and used to refine
the boundaries of the endocardium. In the polar image the edge information is
extracted by applying the Canny edge detection [Cann 86]. To extract minor edges,
a Gaussian smoothing with a standard deviation of σ = 2.5 is applied, as depicted
in Figure 4.5 (b). The lower bound of the hysteresis thresholding was set to 10 % of
the maximum and the upper bound to 20 %.
In addition, the scar probability is evaluated. Therefore, the mean µbp ∈ R and
the standard deviation σbp ∈ R of the blood pool intensity are estimated. In the
next step, a scar threshold θst ∈ R is defined as θst = µbp + σbp . All pixels that are
greater then θst are quantified as potential scar candidates. Knowing the position of
66 Left Ventricle Segmentation in 3-D LGE-MRI
0
r [pixel]
54
0◦ 90◦ 180◦ 270◦ 360◦
ρ
(a) Polar image with transformed mask in polar coordinates (red)
0
r [pixel]
54
0◦ 90◦ 180◦ 270◦ 360◦
ρ
(b) Edge image using Canny edge detector
0
r [pixel]
54
0◦ 90◦ 180◦ 270◦ 360◦
ρ
(c) Scar map
0
r [pixel]
54
0◦ 90◦ 180◦ 270◦ 360◦
ρ
(d) Cost array for the MCP
0
r [pixel]
54
0◦ 90◦ 180◦ 270◦ 360◦
ρ
(e) Cost array with the found contour from MCP
0
r [pixel]
54
0◦ 90◦ 180◦ 270◦ 360◦
ρ
(f ) Final result of the endocardium in the polar image
Figure 4.5: Filter based endocardial contour refinement in polar space. (a) Polar
image and the contour points in polar coordinates (red). (b) Edge information from
Canny edge detector with a Gaussian smoothing of σ = 2.5. (c) Scar map. (d) Cost
array for the MCP search, combining the scar map and the edge image. (e) Cost
array with the found contour from the MCP search. (d) Final result of the MCP in
the polar image.
4.3 Automatic Left Ventricle Segmentation 67
Figure 4.6: Final steps of endocardial refinement. (a) Contour points after conver-
sion from polar to Cartesian coordinates. (b) Convex hull of the endocardial outline
in (a).
the scar candidates, all pixels with increasing radius are labeled with 1, resulting in
a scar map, as visualized in Figure 4.5 (c). The edge image and the scar map are
then combined, resulting in a cost array K ∈ Rr×ρ , see Figure 4.5 (d). The cost
array K is used to initialize a minimal cost path (MCP) search [Dijk 59]. Six equally
distributed contour points are selected from v li to initialize the MCP search, where
the index l corresponds to the current slice. The initialization points are visualized in
Figure 4.5 (d). The number of initialization points for the MCP is chosen heuristically.
The MCP finds the distance weighted minimal cost path through the cost array K.
The cost path is calculated as the sum of the costs at each point of the path, where
edge pixels have 0 costs and non-edge pixels have costs of 1. The result of the MCP
search is visualized in Figure 4.5 (e) and (f).
After the refinement, the contour points v li are converted back to Cartesian coor-
dinates, as illustrated in Figure 4.6 (a). As the result is frayed and papillary muscles
close to the endocardial border may be included, the convex hull is calculated for the
estimated points. This assumption is based on the fact that the left ventricle’s cavity
is convex in the short axis orientation. The final endocardial contour V final is derived
from the smallest convex polygon of the contour points v li , as shown in Figure 4.6 (b).
After the first contour V lfinal ∈ {v li , ..., v lN } is refined, the result is used for further
improvement of the silhouettes in the lower and upper pseudo SA slices, till the base
and the apex are reached. The refinement in basal and apical direction is performed
in polar space for the same reasons as described above. The distance between the
previously found contour points V l−1 final and the silhouette from the transformed atlas
l
mask V final of the next slice is calculated. If the distance between two points is larger
than a margin θdist ∈ R, the contour point of the transformed mask is not considered
for the further improvement,
(
v l−1
i if ||v l−1
i − v li || < θdist
v li = . (4.13)
{} otherwise
68 Left Ventricle Segmentation in 3-D LGE-MRI
In the next step, the set of contour points from the previous slice V l−1 final and the
0
remaining set of contour points from the current slice V l are merged, V l = V l−1 final ∪
l
V and afterwards sub-sampled with a sampling rate of ξ ∈ Z. The sub-sampling
is performed in order to obtain a smoother contour and to remove single outliers.
0
Afterwards, additional refinement steps for the endocardial contour V l approximation
are performed. The polar image is extended to the left (ρ < 0◦ ) and right (ρ >
360◦ ) with γ ∈ [30◦ , 60◦ , 90◦ ], where γ defines the extension of the polar image.
From these enlarged polar images, the edge information is extracted using the Canny
edge detector with a Gaussian smoothing of σ = 3 to extract more dominant edges.
The Gaussian smoothing parameter σ was chosen heuristically. In addition, the
0
sub-sampled silhouette points V l are extended to the left and the right with the
corresponding γ. The enlargement of the polar image is done to avoid errors at the
border of the edge image, in particular if papillary muscles or myocardial infarctions
are present.
Furthermore, the scar map is calculated by applying θst . The cost array is derived
from the scar map combined with the edge image. As before, six equally distributed
0
points are selected from the extended contours V l for the initialization of the MCP
search, for all three cases with respect to γ. First, the shortest path for each of the
three cases is chosen. Second, the length of the path is normalized with respect to the
corresponding enlargement γ. The final contour V lfinal is picked from the results by
choosing the solution with the minimal costs. This outline is then transferred back
to Cartesian coordinates. The convex hull is computed for the refined points V lfinal .
A round shape of the LV in SA view is assumed [McLe 15, Duan 99]. For the
inter-slice smoothness criteria, the shape of the previous contour to the newly found
contour is compared. If the difference of the two contours is bigger than a certain
margin θdiff ∈ R, the previous contour is considered. Depending on the apical or basal
direction, erosion is applied to the contour, as the convex hull slightly enlarges the
contour and in apical direction the radius is decreasing. For further improvements
different constraints are considered depending on the apical or basal direction.
For the apical direction the center cpre of the final contour and the radius rpre
of the semi-axis from the previous slice are examined. As a high resolution is given
through the 3-D acquisition technique, the center of the new curve cnew has to be
within a certain distance θc ∈ R to the center of the previous curve cpre . To be
precise, a shift between two pseudo SA slices is not possible, as they are computed
from the same 3-D volume with isotropic voxel size. Furthermore, the radius rnew
should always be equal or smaller to the radius rpre from the previous refined slice
for the apical direction. If one of these constraints is not met, the previous contour
is taken and erosion is applied to achieve a slightly smaller radius,
(
V lfinal if ||cpre − cnew || < θc ∧ rnew ≤ rpre
V lfinal = l−1
, (4.14)
E(V final ) otherwise
where E denotes binary erosion. The apical iteration ends if the radius rnew of the
semi-axis falls below a threshold θr ∈ R. See Figure 4.7 (b) for an illustration of the
apical refinement direction.
Similar conditions are applied for the basal direction. Here, the center cnew of
the new curve, the center cpre of the previous slice as well as the previous contour
4.3 Automatic Left Ventricle Segmentation 69
Figure 4.7: (a) Refinement for the basal direction, considering the shape and center
of the previous contours to guarantee for inter-slice smoothness. (b) Refinement for
the apical direction considering the radius, shape, and center of the previous contour
to garantee for inter-slice smoothness.
shape and size are considered. If the distance between the centers is greater than a
threshold θc , the endocardial contour is approximated as follows,
(
Vl if ||cpre − cnew || < θc ∧ rnew ≤ rpre
V lfinal = l−1
. (4.15)
V final otherwise
The basal iteration either stops if the maximum point from the transformed mask
is reached or the difference of the areas of the previous contour V l−1 final and the newly
l
found contour V final is greater than a threshold θbmax ∈ R. In this case, the outflow
tract is reached. See Figure 4.7 (a) for an illustration of the basal refinement direction.
All the thresholds for the contour refinement are summarized in Table 4.2.
As for the filter based approach, the learning based method starts with the mid-slice
of the LV, which corresponds to the center of mass of the transformed mask M .
Figure 4.8 (a) depicts the estimated SA view with the contour of the transformed
mask M . Potential boundary points are extracted by circular ray casting, using the
set of contour points V l as initialization, see Figure 4.8 (b).
The potential boundary candidates are classified using a trained RF classifier. The
performance of any classifier is limited to the discriminative power of the features used
70 Left Ventricle Segmentation in 3-D LGE-MRI
Description Symbol
Scar threshold θst
Distance between contour points θdist
Sub-sampling rate ξ
Extension of polar image γ
Difference between areas θdiff
Difference between centers θc
Apex reached epicard θr
Base reached θbmax
Enlargement for epicard θepi
0 .8
0 .6
0 .4
0 .2
0 .0
(c) Boundary costs cendo (d) Boundary costs cendo
Figure 4.8: (a) Estimated SA view using PCA with the contour of the transformed
mask M in red. (b) Potential boundary candidates extracted by circular ray casting
and using the contour of the transformed mask as an initialization. (c) Boundary
costs cendo overlaid on the SA view. (d) Boundary costs cendo as cost array in Cartesian
coordinates.
4.3 Automatic Left Ventricle Segmentation 71
for training. The RF is trained using 16 steerable features, based on local intensity
and gradient [Zhen 08]. Steerable features are used because they are computationally
efficient and can capture orientation and scale. For a given boundary candidate
p = (x, y)T with the intensity √ I and
√ the gradient 5I(p) =pg = (gxp , gy )T , the following
features are extracted:
p I, I, 3 I, I 2 , I 3 , log I, ||g||, ||g||, 3 ||g||, ||g||2 , ||g||3 ,
log ||g||, gx , gy , gx 2 + gy 2 , 52 I(p). Therefore, the polar image is calculated and
the features are extracted in polar space.
The training of the RF is based on gold standard annotations from which positive
and negative samples are extracted. For the training, pathologic as well as healthy
subjects were used to generate a broad range of training data. The RF classifier
assigns each boundary candidate a probability p ∈ [0, 1]. The classification result
is interpreted as costs cendo , where cendo = 1 − p, see Figure 4.8 (c) and (d) as an
example. However, the boundary costs are not sufficient to accurately detect the
endocardial contour. Therefore, an additional scar exclusion step is added. As for
the filter based approach, a scar threshold θst is defined based on the mean intensity
µbp and standard deviation σbp of the blood pool. Having the scar threshold θst , a
scar map is derived, as illustrated in Figure 4.9 (c).
The scar map is impaired with the boundary costs, resulting in a cost array, see
Figure 4.9 (e). To find the final endocardial contour dynamic programming is used
to compute the optimal boundary from the left to the right hand side of the polar
image. Hence, a MCP search is used. The cost path is calculated as the sum of the
costs for each move and weighted by the length of the path. In Figure 4.9 (f) the
result of the MCP search is depicted.
After the contour is refined in the mid slice, this information is used for the suc-
ceeding slices in apical and basal direction. Here, the same steps are applied, the
possible boundary point extraction using circular ray casting, the boundary proba-
bility estimation using the trained RF classifier, the scar map generation, and the
MCP search. Furthermore, similar inter-slice smoothness constraints are applied as
for the filter based approach.
The contour extraction of the epicardium is performed in polar space for the same
reason as mentioned for the endocardial refinement, see Section 4.3.3. The edge infor-
mation is extracted using the Canny edge detector [Cann 86]. The standard deviation
σ of the Gaussian filter kernel is set to σ = 2, where σ is chosen heuristically. As the
epicardium has to be greater then the endocardium, the radius of the endocardial
contour is enlarged by θepi ∈ R, as depicted in Figure 4.10 (a).
In addition, all edges that fall within the endocardial contour are erased from the
edge image, as visualized in Figure 4.10 (b). Having the enlarged endocardial contour,
72 Left Ventricle Segmentation in 3-D LGE-MRI
0
r [pixel]
54
0◦ 00◦ 180◦ 270◦ 360◦
ρ
(a) Polar image
0 scar
r [pixel]
54 healthy
0◦ 90◦ 180◦ 270◦ 360◦
ρ
(b) Scar classification
0 1
r [pixel]
54 0
0◦ 90◦ 180◦ 270◦ 360◦
ρ
(c) Scar map
0 1
r [pixel]
0.5
54 0
0◦ 00◦ 180◦ 270◦ 360◦
ρ
(d) Boundary costs cendo in polar space
0 1
r [pixel]
0.5
54 0
0◦ 00◦ 180◦ 270◦ 360◦
ρ
(e) Cost array derived from the scar map and the boundary costs
0
r [pixel]
54
0◦ 00◦ 180◦ 270◦ 360◦
ρ
(f ) Result of the MCP in polar space
Figure 4.9: (a) Polar image. (b) Scar probability. (c) Scar map derived from the
scar threshold θst . (d) Boundary costs cendo in polar space. (e) Cost array derived
from the combination of the scar map with the bundary costs array. (f) Final result
of the endocardial boundary estimation using a MCP search.
4.3 Automatic Left Ventricle Segmentation 73
r [pixel]
54
0◦ 90◦ 180◦ 270◦ 360◦
ρ
(a) Epicardial contour enlarged by θepi
0
r [pixel]
54
0◦ 90◦ 180◦ 270◦ 360◦
ρ
(b) Epicardial edges only
Figure 4.10: (a) The final endocardial boundary is visualized in red and the enlarged
epicardial contour in yellow. (b) Canny edge image showing only the edges that do
not fall within the endocardial contour.
the closest edge with increased radius is searched for in the epicardial edge image. A
polynomial is fitted through the found points to remove outliers. In the next step,
the extracted contour points are transformed back to Cartesian coordinates. As the
result might be frayed, the convex hull is estimated to smooth the contour. Due to
the smoothing operation, the epicardial contour is slightly enlarged, hence, erosion is
applied to shrink the outline. The result of the endocardial and epicardial contours
are shown in Figure 4.11.
For the basal direction, the iterations of the epicardial contour refinement stop
with the refinement of the endocardial contour, i. e., when the left ventricular outflow
tract (LVOT) is reached. For the apical direction, there is no endocardial contour
anymore, as the endocardial contour refinement stops when the semi-axis falls below
a certain threshold θr . Hence, the previously estimated epicardial contour is used
for the refinement with decreasing radius. The epicardial contour refinement for the
apex stops, when the semi-axis of the epicardium falls below a certain threshold θrepi .
As for the learning based endocardial contour detection, which is described in Sec-
tion 4.3.3, a RF classifier is used to determine possible boundary candidates. The
RF is trained with the same 16 steerable features as for the endocardial boundary
estimation. For the training, ground truth annotations of the epicardial boundary
are used, where positive as well as negative samples are extracted. To extract pos-
sible epicardial boundary candidates, the previously found endocardial boundary is
used and the radius is enlarged by θepi . From the enlarged contour, the boundary
candidates are extracted using circular ray casting, which is performed in polar space.
The boundary candidates are classified using the trained RF, resulting in a boundary
probability p ∈ [0, 1]. The result of the epicardial boundary classification is used as a
cost array for the MCP search. The MCP search finds the distance weighted minimal
74 Left Ventricle Segmentation in 3-D LGE-MRI
Figure 4.11: Final result of the endocardial (red) and epicardial (yellow) contour
in Cartesian coordinates.
path through the cost array. The result of the MCP is transferred back to Cartesian
coordinates and the smallest convex polygon is estimated to achieve a smooth looking
contour.
The refinement of the basal and apical direction ends with the same conditions
as for the filter based approach, which is described in the previous Section 4.3.4.
Figure 4.12: Image of the human heart, where the right and left ventricle is open.
The papillary muscles are connected to the mitral valve via the chordae tendineae.
The image "Heart normal short axis left ventricle view with papillary muscles" by
Patrick J. Lynch is licensed under CC BY 2.5.
(a) 2-D papillary muscles (b) 3-D papillary muscles (c) 3-D surface mesh
Figure 4.13: (a) Final result showing the endocardial contour in red and the papil-
lary muscles in cyan. (b) Papillary muscles visualized as 3-D surface mesh in cyan and
the endocardial surface in red. (c) Final segmentation result of the endocardial (red)
and epicardial (yellow) surface, with a myocardial infarction visualized in purple.
rithm is used to obtain the iso-surfaces of the contours [Lore 87]. The output of the
algorithm is a triangular mesh consisting of a set of vertices and connected faces. The
extracted vertices and faces, as well as face normals are then saved in standard sur-
face mesh file formats. Figure 4.13 (c) illustrated an example of a 3-D surface mesh,
where the endocardial surface is red, the epicardial surface yellow, and a myocardial
infarction is visualized in purple.
76 Left Ventricle Segmentation in 3-D LGE-MRI
Figure 4.14: Semi-automatic segmentation pipeline for the left ventricle segmenta-
tion. The first image from the left shows the 3-D volume as input. In the next step,
single slices are segmented using the smart brush functionality. Third, the control
points of the contours are extracted. Fourth, the 2-D and 3-D normal vectors are
computed for the HRBF interpolation. In the final image, the interpolated surface is
visualized.
Figure 4.15: (a) Initialization of the smart brush in red and the smart brush is
propagated which is illustrated as yellow circle. (b) Segmented area under the brush
using adaptive thresholding and propagation checking.
In the next step, the control point (CP)s are computed from the contours adaptively
according to the shape of the object.
The contour is sampled equidistantly with a predefined sampling size ξ ∈ Z. The
number of control points ne ∈ N is based on the contour length k ∈ R+
0 and computed
as follows
k
ne = b c . (4.16)
ξ
Furthermore, nc ∈ N convexity defect points, where the contour has the maximum
distance to its convex hull are added, see the blue points in Figure 4.16 (a). To
increase the accuracy of the 3-D interpolation for complex objects, the number of
CPs is increased at rough areas. Therefore, the local curvature κ ∈ R is checked for
all CPs and additional points are added in case of roughness. The curvature κ is
dependent on the derivatives of the curve k ∈ R2 with k(t) = (x(t), y(t))T ,
|x0 y 00 − y 0 x00 |
κ= , (4.17)
(x02 + y 02 )3/2
where the primes refer to derivatives dtd with respect to the parameter t. To compare
curvature values, a reference quantity r ∈ R (global roughness) is defined which is
the ratio of the convex hull area of the curve Ah and the curve area Ac [Abbe 17],
Ac
r= . (4.18)
Ah
New CPs are added at a certain distance to the investigated CP, if the criterion
κ
> θκ (4.19)
r
is fulfilled, where the threshold θκ ∈ R is obtained heuristically. The number of
additional CPs due to curvature is denoted as nκ . The total number of CPs is
4.4 Semi-Automatic Left Ventricle Segmentation 79
Figure 4.16: Control point extraction. (a) A rough surface with initial equidistant
points in red and convexity defect points in blue. (b) A rough surface with increased
number of points in green.
Figure 4.17: Two orthogonal planes are segmented in red and the resulting inter-
section points are depicted in yellow.
of the 3-D interpolation. The calculation of the intersection points and their normal
vectors is explained in the following section.
Contour Intersection
Since the volumetric 3-D image is segmented in arbitrary 2-D slices, intersections
between segmented slices from different orientations occur. Suppose that closed con-
tours are extracted from the intersecting slices, the intersection should result in two
3-D points. The computation of these intersection points is performed iteratively.
The iterations are over all non-repetitive 2-permutations of N segmented slices and
comprise of two steps.
First, the existence of intersections is checked. In case of an intersection, the cor-
responding points are extracted. In some complex objects, i. e., non-convex shapes,
more than two intersection points or a set of neighboring points are extracted. Fig-
ure 4.18 shows two possible cases for extracted intersection points. In case of multiple
intersection candidates, classification is applied in order to distinguish between the
different groups of points. In the next step, the average of each group is taken as final
3-D intersection point. The classification of the point groups is described in more
detail in the following section.
Having found the intersection point candidates at each junction, the next step is to
merge close intersection points in order to decrease the redundancy. To obtain the
neighboring points, a nearest neighbor graph within a given radius rp is applied to the
set of intersection point candidates. According to the size and the complexity of the
desired object, the user can change the search radius rp for the neighboring points.
The algorithm does not merge close points from parallel planes, otherwise information
loss would occur. The merging of the 3-D points is simply done by averaging, resulting
in one 3-D point. The merging of the normal vectors n is performed such that the
4.4 Semi-Automatic Left Ventricle Segmentation 81
m m m m
!T
1 X 1 X 1 X 1 X
p= pi = xi , yi, zi , (4.20)
m i=1 m i=1 m i=1 m i=1
where pi refers to the intersection point candidates in one neighborhood and m is the
number of candidates within the search radius rp . To have a normal unit vector the
merging is done as
m m m
!T
1 X 1 X 1 X
n= ni,x , ni,y , ni,z , (4.21)
N (ni,x ) i=1
N (ni,y ) i=1
N (ni,z ) i=1
n
nnormalized = , (4.22)
knk
where operator N counts the number of valid elements, i. e., the intersection points
that have an in-plane normal for this dimension. In Figure 4.19, all the control points
extracted from the different image planes are shown, where the control points with
a 2-D normal vector are visualized in green and all control points with a 3-D normal
vector are shown in blue.
Figure 4.19: Control points extracted from three different orientations, where N
points have a 2-D normal vector in green and M points with 3-D normal vector in
blue.
where α ∈ R is a weighting factor for each control point. To make sure that the
equation is always solvable a low-degree polynomial g (c) is added
N
X
f (c) = αi ϕ (kc − pi k) + g (c) . (4.25)
i=1
However, this simple RBF formulation requires the definition of inside and outside
values. To address this issue, Hermite data is incorporated into the RBF, which
directly use derivatives. This method ensures the existence of a non-null implicit
surface without the need of additional information [Mace 11]. Using the first order
Hermite interpolation in combination with RBF, the scalar field can be formulated
as follows:
XN
f (c) = αi ϕ (kc − pi k) − β i ∇ϕ (kc − pi k) + g (c) , (4.26)
i=1
scalar field f consists of two components f = f 2D + f 3D . The scalar field f 2D for the
2-D normal vectors is formulated as
N
T
X
2D
f (c) = α2D 2D
i ϕ (kc − pi k) − β i h2D
i (∇ϕ (kc − pi k)) + g (c) , (4.27)
i=1
where g(c) is a low-degree polynomial, h2D i (c) is a function that selects the 2-D
gradient direction that is available for control point pi , and α2D
i ∈ R, β 2D
i ∈ R2 are
3D
the RBF coefficients. The scalar field f for the 3-D normal vectors is formulated
accordingly,
NX
+M
T
f 3D (c) = α3D 3D
i ϕ (kc − pi k) − β i ∇ϕ (kc − pi k) + g (c) , (4.28)
i=N +1
have to be fulfilled [Ijir 13]. These constraints yield a linear system of equations
represented in the matrix form as
O T1 ··· O TM O TM +1 ··· O TM +N
0 s 0
O1 G1,1 ··· G1,M G1,M +1 ··· G1,M +N w1 b1
.. .. .. .. .. .. .. .. ..
. . . . . . .
.
.
OM GM,1 · · · GM,M GM,M +1 ··· GM,M +N wM = bM ,
O M +1 GM +1,1
· · · GM +1,M GM +1,M +1 ··· GM +1,M +N wM +1 bM +1
(4.36)
. .. .. .. .. .. .. . .
.. . . . . . . .. ..
O M +N GM +N,1 · · · GM +N,M GM +N,M +1 ··· GM +N,M +N wM +N bM +N
| {z } | {z } | {z }
=D =Y =B
84 Left Ventricle Segmentation in 3-D LGE-MRI
where different color in the matrix implies the points and the corresponding normal
vectors with different dimensionality (3-D blue, 2-D green, mixed purple). The linear
systems of equations can be also written as DY = B, with D ∈ RM +N ×M +N , Y ∈
RM +N and B ∈ RM +N .
The blue block describes the constraints on the 3-D variables α3D 3D
i , βi derived
from the 3-D constraints and orthogonality conditions Equations (4.29), (4.31), (4.33)
and (4.35). Thus, the matrices Gi,j , O i and the vectors c, wi , and bi are defined as:
T
ϕ(kpi − pj k) −∇ϕ(kpi − pj k)T pi 1
Gi,j = , Oi =
∇ϕ(kpi − pj k) −Hϕ(kpi − pj k) E 0
3D (4.37)
a αi 0
s= , wi = 3D , bi = ,
b βi ni
where E ∈ R3×3 is a unit matrix and Hϕ ∈ R3×3 is the Hessian matrix of the kernel ϕ,
which arises due to the normal constraint Equation (4.31) applied to ∇ϕ. The green
block describes the constraints on the 2-D variables αi2D , βi2D derived from the 2-D
constraints and orthogonality conditions Equations (4.29), (4.30), (4.32) and (4.34).
Thus, the matrices GM +i,M +j and O M +i and the vectors wM +i and cM +i are defined
as:
(∇ϕ(kpi − pj k))T
ϕ(kpi − pj k) −h2D
i
GM +i,M +j = 2D ,
hi (∇ϕ (kpi − pj k)) −h2Di ∇T h2D
i (∇ϕ (kpi − pj k))
2D T
h pi 1
O M +i = i (4.38)
E 0
2D
αi 0
wM +i = 2D , bM +i = .
βi n2D
i
The mixed blocks are defined analogously. There is always a unique solution to the
system of equations, if the points pi are pairwise distinct [Braz 10, Ijir 13].
Considering the basis function of the tri-harmonic kernel ϕ(t) = t3 , the gradient
and the Hessian matrix of the kernel ϕ is denoted as follows:
∇ϕ(t) = 3 ktk t
(
0 if ktk = 0 (4.39)
Hϕ = 3ttT ,
ktk
+ 3 ktk E k otherwise
The function ψ has many other level sets, in addition to υ, while only υ has a meaning
for the segmentation and not for any other level sets of ψ. A very commonly chosen
function ψ is the signed distance to the front υ(0) given as
−d(x, y, 0) if (x, y) inside the front
ψ(x, y, 0) = 0 if (x, y) on the front . (4.42)
d(x, y, 0) if (x, y) outside the front
The level set method segments the surface iteratively. In the first step, the front υ(0)
is initialized at a certain position. The second step is to compute ψ(x, y, 0) and then
iterate over until convergence.
(a) Cohort 1 SA view (b) Cohort 1 LA view (c) Cohort 2 SA view (d) Cohort 2 LA view
Figure 4.21: (a) and (b) shows one example data set of the first cohort in a SA and
LA axis orientation. (c) and (d) depicts an example data set of the second cohort.
Here the voxel spacing is not isotropic and no additional fat suppression was applied.
This results in further enhancements in the apical region of the left ventricle, see (d).
Table 4.3: Parameters for the left ventricle segmentation. These parameters are
chosen heuristically.
4.5.2 Evaluation
Gold standard annotations of the LV endocardium and epicardium are provided by
two clinical experts and are performed using various open source segmentation tools
(Slicer [Fedo 12], Seg3D [CIBC 15], and MITK [Nold 13]). The three tools are equiv-
alent and have no influence on the segmentation result, as always the brush func-
tionality is used. The observers are asked to segment the endocardial and epicardial
contour.
For the automatic segmentation of the endo- and epicardium, several parameters
have to be chosen heuristically. A summary of all the parameters is provided in
Table 4.3. The influence of the parameter values is discussed in Section 4.5.3.
Metrics
Given the gold standard annotation, the segmentation is evaluated using two different
measures. First, the Dice coefficient (DC) as a quantitative score of the segmentation
quality is evaluated, as it measures the proportion of the true positives in the seg-
mentation. Dice scores range from 0 to 1, with 1 corresponding to a perfect overlap.
For the definition of the Dice coefficient please refer to Equation (3.3). For the DC,
the whole 3-D volume is considered.
The second evaluation method is the mean surface distance (MSD) between the
surface voxel of the binary mask A and their nearest surface voxel of binary object
B, averaged over all contour points and all slices of the volume.
88 Left Ventricle Segmentation in 3-D LGE-MRI
Figure 4.22: (a) Example of the non-smoothed gold standard annotation in the
SA orientation. It can be seen that the contours look frayed. (b) Smoothed gold
standard orientation in the SA orientation.
Assessment
The two different measures are applied to the whole volume, differentiating the endo-
and epicardium. Furthermore, the inter-observer variability between the observers is
investigated. In addition, the volume is separated in three parts, the base, mid-cavity,
and apex, by dividing the gold standard annotation into equal thirds, perpendicular
to its long axis [Ma 12].
Smoothing
The gold standard annotation is done in the sagittal, coronal, and axial plane, there-
fore, the results look frayed in the SA orientation, see Figure 4.22 (a) for an example.
The data was already annotated by the physician, therefore all the annotations are
performed in the standard image planes. However, to overcome this issue with the
frayed contours, the gold standard annotation is post-processed by applying the con-
vex hull to the contour points for every slice in the SA orientation. To be more
precise, the convex hull is estimated in 2-D for every slice without considering the
overall 3-D shape. An example of the smoothed contour is shown in Figure 4.22 (b).
Hence, the DC and MSD are computed for the gold standard annotations with and
without smoothing and also the difference between the gold standard annotation with
and without smoothing is evaluated.
4.5.3 Results
First, the results for the filter based segmentation are presented. Then the effect of
the smoothing is evaluated and the parameter variability is investigated. Afterwards,
the results for the learning based segmentation are shown. Furthermore, the filter
based and learning based results are compared against each other. Finally, the semi-
automatic segmentation approach using the smart brush and the adopted HRBF
interpolation is evaluated.
4.5 Evaluation and Results 89
Endo Epi
Cohort 2 Cohort 1 Description Mean ± Std Min Max Mean ± Std Min Max
Mean 0.84 ± 0.06 0.73 0.93 0.85 ± 0.05 0.73 0.90
Observer 1 0.84 ± 0.06 0.73 0.91 0.84 ± 0.05 0.73 0.89
Observer 2 0.85 ± 0.06 0.74 0.93 0.85 ± 0.04 0.76 0.90
Inter-Observer 0.94 ± 0.02 0.92 0.98 0.93 ± 0.02 0.95 0.95
Mean 0.83 ± 0.03 0.76 0.89 0.78 ± 0.04 0.68 0.85
Observer 1 0.83 ± 0.03 0.77 0.89 0.78 ± 0.03 0.70 0.82
Observer 2 0.83 ± 0.03 0.76 0.89 0.78 ± 0.05 0.68 0.85
Inter-Observer 0.90 ± 0.04 0.81 0.98 0.87 ± 0.05 0.75 0.96
(a) Dice coefficient without smoothing
Endo Epi
Description Mean ± Std Min Max Mean ± Std Min Max
Cohort 2 Cohort 1
Table 4.4: Quantitative results of filter based 3-D LV segmentation. The results are
shown separately for the endocardial (Endo) and epicardial (Epi) contour. Further-
more, it is distinguished between the two different clinical cohorts. (a) DC without
smoothing of the gold standard annotation. (b) DC with smoothing of the gold
standard annotation.
First, the gold standard without smoothing is considered. The automatic seg-
mentation of the endocardium results in an overlap to the gold standard annotation
of 0.83 ± 0.04 considering the DC. The best segmentation result has a DC of 0.93
and the worst a DC of 0.73 because of a huge myocardial scarring. For the epicard,
an overall DC of 0.80 ± 0.05 is achieved. The best segmentation of the epicardium
yields a DC of 0.90 and the worst a DC of 0.68. As two gold standard annotations
are available, the inter-observer variability is addressed. The inter-observer variabil-
ity between the two observers results in a DC of 0.92 ± 0.04. The best inter-observer
variability results in a DC of 0.98 and the worst in a DC of 0.81. In Table 4.4 (a) a
distinction between the two different clinical cohorts is made. It can be seen that the
results for the endocardial contour extraction are similar with a DC of around 0.83.
However, the results for the epicard are worse for cohort two. This correlates with a
reduced inter-observer variability of 0.87 for the second cohort’s epicardium.
90 Left Ventricle Segmentation in 3-D LGE-MRI
1.0 1.0
Dice Coefficient
Dice Coefficient
0.8 0.8
0.6 0.6
All All
0.4 Base 0.4 Base
Mid-Cavity Mid-Cavity
0.2 0.2
Apex Apex
0.0 0.0
0 20 40 60 80 100 0 20 40 60 80 100
Percentage [%] Percentage [%]
Figure 4.23: (a) The gold standard without smoothing is used. Considering the
endocardium and epicardium together, an average DC of 0.82 ± 0.07 is achieved.
Furthermore, the LV is divided into three parts, the base, mid-cavity, and apex.
(b) The gold standard with smoothing is used. Considering the endocardium and
epicardium together, an average DC of 0.82 ± 0.07 is achieved. Furthermore, the LV
is divided into three parts, the base, mid-cavity, and apex.
1.0
0.9
Dice Coefficient
0.8
0.7
0.6
Endocard
Epicard
0.5
1 5 10 15 20 25 30
Sequence Number
Figure 4.24: Individual Dice coefficient for 3-D filter based segmentation for each of
the 30 data sets for the endocard and epicard, respectively. The DC for the endocard
is sorted in increasing order. The DC for the epicard is sorted according to the
endocard.
required for the mesh generation. Furthermore, the results in the mid-cavity are very
convincing and are adequate for the scar quantification. However, the segmentation
of the apex for some cases is not sufficient especially for the second cohort. The
decreased quality is due to the hyper-enhancements in the apex as for the second
cohort no fat suppression is applied.
Now, the results for the mean surface distance are reported. First, the gold
standard without smoothing is considered. Using the MSD, the endocard has a
mean distance of 2.80 mm ± 0.80 mm, with a minimum of 1.14 mm and a maximum of
4.85 mm. The epicard has a mean distance of 4.17 mm ± 1.15 mm, with a minimum
average of 1.82 mm and a maximum average of 6.75 mm. The inter-observer variability
between the two observers results in a mean surface distance of 0.91 mm ± 0.04 mm,
with a minimum of 0.81 mm, and a maximum of 0.98 mm. In Table 4.5 (a) it is
differentiated between the two different clinical cohorts. The mean surface distance
of the epicard for the second cohort is worse compared to the first cohort. This
observation correlates with the MSD of the inter-observer variability for the epicard.
For the gold standard annotation with smoothing, the endocard has a mean
distance of 2.75 mm ± 0.73 mm, with a minimum of 1.49 mm and a maximum of
4.48 mm. The epicard has a mean distance of 4.29 mm ± 1.18 mm, with a mini-
mum average of 1.90 mm and a maximum of 6.72 mm. The inter-observer variability
between the two observers results in a MSD of 1.35 mm ± 0.73 mm, with a min-
imum mean distance of 0.30 mm, and a maximum mean distance of 3.05 mm. In
Table 4.5 (b) there is a distinction for the two different clinical cohorts. Here, the
same observations can be seen as for the DC with smoothing.
92 Left Ventricle Segmentation in 3-D LGE-MRI
Table 4.5: Quantitative results of filter based 3-D LV segmentation, using the MSD.
The results are shown separately for the endocardial (Endo) and epicardial (Epi)
contour in mm. Furthermore, it is distinguished between the two different clinical
cohorts. (a) MSD without smoothing of the gold standard annotation. (b) MSD with
smoothing of the gold standard annotation.
Qualitative Results
Figure 4.25 shows an example for a qualitative evaluation for one clinical data set
of the first cohort. The first row shows the pseudo SA slices from basal to apical
direction without any contours. The second row depicts the gold standard annotation
without smoothing from the physician, where the endocardium is marked in orange
and the epicardium in green. The third row illustrates the gold standard annotation
with smoothing, where the endocardium is marked in orange and the epicardium
in green. The fourth row delineates the filter based segmentation result, where the
endocardial contour is red and the epicardial contour is yellow. It can be seen that the
presented result matches well with the smoothed gold standard annotation. However,
the manual annotation looks frayed. Furthermore, it can be observed that the biggest
variance of the annotations is in the apex, which is also shown in Figure 4.23 regarding
the DC.
4.5 Evaluation and Results 93
Figure 4.25: Comparison of the segmentation result for sequence number seven from
the first cohort. The first row shows the pseudo SA slices from basal to apical direction
without any contours. The second row depicts the gold standard annotation from the
physician without smoothing, where the endocardium is marked in orange and the
epicardium in green. The third row depicts the smoothed gold standard annotation,
where the endocardium is marked in orange and the epicardium in green. The fourth
row delineates the filter based result of the presented segmentation algorithm, where
the endocardial contour is red and the epicardial contour is yellow.
94 Left Ventricle Segmentation in 3-D LGE-MRI
Physician
Expert
(a) Endocardium
Physician
Expert
(b) Epicardium
Figure 4.26: Comparison of the segmentation result for the two observers. The first
row depicts the smoothed gold standard annotation from the endocardium, where
the lighter orange contour is from the physician and the darker contour from the
clinical expert. The second row shows the smoothed gold standard annotation of the
epicardium, where the light green contour is from the physician and the dark green
from the clinical expert.
Figure 4.26 compares the two different gold standard annotations obtained from the
clinical experts, from base to apex. In the first row, the smoothed gold standard
annotations from the endocardium are shown. In the second row, the smoothed
gold standard annotations from the epicardium are shown. In this data set, a huge
myocardial scarring is present. It can be seen that the contours differ especially in
the apex and at the left ventricular outflow tract. These regions also show the largest
differences compared to the fully automatic segmentation.
In addition to the inter-observer variability, the Dice coefficient and mean sur-
face distance between the smoothed gold standard annotation and non-smoothed
annotations is evaluated. The endocard has a mean DC of 0.96 ± 0.01. The best
segmentation overlap has a DC of 0.97 and the worst a DC of 0.89. The epicard has
a mean DC of 0.96 ± 0.02. The best match yields a DC of 0.98 and the worst a DC
of 0.92.
Using the mean surface distance, the endocard has a mean distance of 0.68 mm ±
0.32 mm, with a minimum of 0.36 mm and a maximum of 2.41 mm. For the epicard, a
mean distance of 0.79 mm ± 0.40 mm is evaluated, with a minimum of 0.33 mm and a
maximum distance of 1.92 mm. The results for the smoothing impact are summarized
in Table 4.6.
4.5 Evaluation and Results 95
1.0 1.0
Dice Coefficient
Dice Coefficient
0.9 0.9
0.8 0.8
0.7 0.7
0.6 0.6
0.5 0.5
θdist = 2 θdist = 4 θdist = 6 ξ=2 ξ=4 ξ=8
1.0 1.0
Dice Coefficient
Dice Coefficient
0.9 0.9
0.8 0.8
0.7 0.7
0.6 0.6
0.5 0.5
θdif f = 20 θdif f = 25 θdif f = 30 θC = 3 θC = 5 θC = 7
1.0 1.0
Dice Coefficient
Dice Coefficient
0.9 0.9
0.8 0.8
0.7 0.7
0.6 0.6
0.5 0.5
θbmax = 178 θbmax = 188 θbmax = 198 θepi = 4 θepi = 6 θepi = 8
Description DC MSD
Endo 0.94 ± 0.02 0.74 ± 0.22
Epi 0.96 ± 0.01 0.88 ± 0.33
Table 4.6: Comparison of the gold standard annotation without smoothing to the
smoothed gold standard annotation using the Dice coefficient (DC) and the mean
surface distance (MSD). The results are shown separately for the endocardial (Endo),
epicardial (Epi) contour.
Parameter Variability
As many parameters are set, as shown in Table 4.3, the sensitivity of these parameters
is evaluated, see Figure 4.27. It can be seen that sensitivity of the distance between
two points θdist , the distance between centers θC , and the base threshold θbmax is low
and the mean value is always in the same range. The sensitivity for the sub-sampling
rate ξ is higher, as the sub-sampled points are the input for the contour refinement.
Furthermore, the sensitivity for the distance between two areas θdiff is higher as this
is an important criteria, if the newly estimated contour is used or the previously
segmented contour. For θepi , it can be seen, that the radius should be enlarged by 6
pixels, otherwise the enlargement is not big enough and thus the contour estimation
using the filter based approach is not as accurate.
As the smoothing has no big influence on the Dice coefficient as shown in Table 4.6
and the results seem more physiologically meaningful compared to the non-smoothed
results as seen in Figure 4.22, only the smoothed gold standard annotations are
considered for the further evaluation of the learning based approach. As the data sets
from the two cohorts differ significantly, only the first cohort is evaluated using the
Endocardium Epicardium
Description T D T D
1
Hendo 70 15 80 15
2
Hendo 60 15 70 15
3
Hendo 50 15 70 15
4
Hendo 60 15 70 15
5
Hendo 60 15 60 15
6
Hendo 50 15 70 15
7
Hendo 60 15 50 15
8
Hendo 50 70 4 30
9
Hendo 60 15 80 15
Table 4.7: Optimized hyper-parameters for 3-D random forest, for each of the
individual folds of the first cohort, for the endocardium and epicardium, respectively.
For the grid-search the following parameter sets are optimized: number of trees T
and maximal tree depth D, for the endocardium and epicardium, respectively.
4.5 Evaluation and Results 97
Table 4.8: Quantitative results of the learning based 3-D LV segmentation using
the DC with smoothing of the gold standard annotations. The results are shown
separately for the endocardial (Endo) and epicardial (Epi) contour.
0.9
0.8
Dice Coefficient
0.7
0.6
0.5
Endocard
0.4
Epicard
0.3
1 2 3 4 5 6 7 8 9
Sequence Number
Figure 4.28: Individual Dice coefficient for 3-D learning based segmentation for
each of the 9 data sets of the first cohort for the endocard and epicard, respectively.
learning based segmentation. The first cohort is chosen, as this acquisition protocol
is the advanced imaging protocol with an isotropic resolution. Furthermore, for the
first cohort fat suppression is applied to remove the enhancements of the pericardial
fat, which could lead to inaccurate segmentation results especially in the apex.
For the evaluation of the first cohort, a leave-one-out cross-validation is performed
for the nine subjects. For the evaluation of the hyper-parameters, a grid-search with
the following parameter sets is evaluated: number of trees T ∈ {50, 60, 70, 80} and
maximal tree depth D ∈ {10, 15, 20, 25, 30}.
The optimal hyper-parameters of the random forest of the endocardium and epi-
cardium are summarized in Table 4.7 for each of the folds.
The learning based segmentation of the left ventricle results in a Dice coefficient
of 0.84 ± 0.07 for the endocardium. The best segmentation results in a DC of 0.91
and the worst in a DC of 0.67. For the epicardium, an overall DC of 0.85 ± 0.07 is
achieved. The best segmentation of the epicardium yields a DC of 0.90 and the worst
a DC of 0.65. The segmentation results for the individual observers are summarized
in Table 4.8. In Figure 4.28 the Dice coefficients for the individual data sets are
represented for the endocard and epicard, respectively.
In addition, also the mean surface distance is evaluated. Using the MSD, the
endocard has a mean distance of 2.96 mm ± 1.25 mm, with a minimum of 1.66 mm and
a maximum of 5.74 mm. The epicard has a mean distance of 3.25 mm ± 1.11 mm, with
a minimum distance of 1.95 mm and a maximum distance of 5.76 mm. In Table 4.9
it is differentiated between the two observers.
98 Left Ventricle Segmentation in 3-D LGE-MRI
Table 4.9: Quantitative results of learning based 3-D LV segmentation, using the
MSD with smoothing of the gold standard annotation. The results are shown sepa-
rately for the endocardial (Endo) and epicardial (Epi) contour in mm.
0.15 0.15
Importance
Importance
0.10 0.10
0.05 0.05
0.00 0.00
q ||g||
q ||g||
log(I)
log(I)
log(||g||)
log(||g||)
I
I
I
I
gx
gx
3
2
3
2
3
gy
gy
5 I(p)
5 I(p)
||g||
||g||
||g||
||g||
√
√
||g||
q||g||
||g||
q||g||
y
y
2 + g2
2 + g2
I
I
3
3
2
2
3
3
gx
gx
q
Feature Feature
Figure 4.29: Feature importance of the random forest classifier for the 3-D
LGE-MRI classification. (a) Features importance for the random forest classifier
trained for the endocardial boundary detection. (b) Feature importance for the ran-
dom forest classifier trained for the epicardial boundary detection. It can be seen,
that the feature importance correlates for both trained random forest classifiers.
Furthermore, the feature importance for the endocard and epicard is evaluated,
see Figure 4.29 (a) and (b), respectively. It can be seen that the order of the features
with respect to their importance is identical for the endocard and epicard. In general,
the gradient features are more important compared to the intensity features. This
can be attributed to the fact of the non-homogeneous intensity distribution of the
LGE-MRI sequences in case of myocardial scarring.
In this section, the results of the filter based approach are compared to the learning
based approach for the first cohort. In Figure 4.30 the filter based results vs. the
learning based segmentation are visualized for the endocardium and epicardium, re-
spectively.
Furthermore, the correlation between the filter based and learning based seg-
mentation is investigate using a scatter plot, as depicted in Figure 4.31 (a) for the
endocard and in Figure 4.31 (b) for the epicard. The Pearson correlation between
the endocardium segmentation results of the two methods results in 0.86 and in 0.84
for the epicard. Therefore, there is a good correlation between the two segmentation
methods.
4.5 Evaluation and Results 99
1.0 1.0
Dice Coefficient
Dice Coefficient
0.9 0.9
0.8 0.8
0.7 0.7
0.6 0.6
0.5 0.5
Filter Based Learning Based Filter Based Learning Based
Dice Coefficient
0.8 0.8
0.6 0.6
Filter Filter
0.4 Learning 0.4 Learning
1 2 3 4 5 6 7 8 9 1 2 3 4 5 6 7 8 9
Sequence Number Sequence Number
Figure 4.30: (a) Comparison of the Dice coefficients between the filter based and
the learning based segmentation for the endocardium, where the blue line represents
the mean Dice coefficient. (b) Comparison of the Dice coefficients between the filter
based and the learning based segmentation for the epicardium, where the blue line
represents the mean Dice coefficient. (c) Individual DC of the endocard for the filter
based approach compared to the learning based approach. (d) Individual DC of the
epicard for the filter based approach compared to the learning based approach.
1.0 1.0
DC of Learning Based
DC of Learning Based
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0.0 0.0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
DC of Filter Based DC of Filter Based
Figure 4.31: Scatter plot of the Dice coefficient of the filter based segmentation
and the learning based segmentation. (a) The Pearson correlation for the endocard
results in 0.86. (b) The Pearson correlation for the epicard results in 0.84.
100 Left Ventricle Segmentation in 3-D LGE-MRI
The qualitative results are compared in Figure 4.32 for sequence number one of the
first cohort. The first row shows the pseudo SA slices from basal to apical direction
without any contours. The second row shows the gold standard annotation from one
clinical expert. The third row delineates the final filter based segmentation result.
The fourth row depicts the results of the learning based segmentation.
As the smoothing has no big influence on the Dice coefficient as shown in Table 4.5
only the smoothed gold standard annotations are considered for the further evalua-
tion of the semi-automatic segmentation based approach. As for the learning based
approach, only the first cohort is used for the evaluation of the semi-automatic LV
segmentation. The 2-D ground truth annotations were used to assess the 2-D seg-
mentation and the complete 3-D ground truth for the 3-D interpolation scheme.
Figure 4.32: Comparison of the segmentation result of the filter based vs. the
learning based approach for sequence number one of the first cohort. The first row
shows the pseudo SA slices from basal to apical direction without any contours.
The second row shows the gold standard annotation from the physician, where the
endocardial contour is orange and the epicardial contour is green. The third row
delineates the final filter based segmentation result, where the endocardial contour
is red and the epicardial contour is yellow. The fourth row depicts the results of the
learning based segmentation, where the endocardial contour is visualized in red and
the epicardial contour in yellow.
102 Left Ventricle Segmentation in 3-D LGE-MRI
1.0
0.8
Dice Coefficient
0.6
0.4
0.2
0.0
1 2 3 4 5 6 7 8 9
Patient Number
Figure 4.33: The evaluation of the 2-D segmentation result using the smart brush
for the first cohort.
Figure 4.34: Smart brush outlier evaluation. (a) The overlaid ground truth shown
in red and the smart brush patch shown as a yellow rectangular. (b) The extracted
patch from the smart brush. (c) The pre-segmented mask which is obtained by
eroding the extracted patch. (d) The segmentation result by using the smart brush
which is different to the ground truth patch due to similar intensity values.
has a MSD of 2.08 mm ± 0.40 mm, 1.39 mm ± 0.38 mm, and 1.14 mm ± 0.25 mm for
1, 3, and 5 slices per orientation, respectively.
The main difference from our proposed A-HRBF method to the HRBF [Ijir 13] is
that we use a combination of 2-D and 3-D gradients based on the extracted contour
of the 2-D segmentation, which makes the interpolation faster. The standard HRBF
method uses the 3-D intensity gradient for their 3-D interpolation. These can lead
to errors, especially in case of ambiguous boundaries, such as the transition between
the left and the right ventricle or areas of myocardial scarring.
Figure 4.36 depicts the qualitative results of the A-HRBF 3-D interpolation scheme
for one example data set, where the result is shown in red and the ground truth in
orange.
In contrast to previous implicit methods for 3-D interpolation [Ijir 13], this method
cannot only be used for high-contrast images, but also for images with high noise level
or other confounding factors due to the independence of intensity information for the
3-D interpolation. The main advantage happens when there is an ambiguous bound-
ary which only an expert can recognize (e. g., between left ventricle and left atrium
104 Left Ventricle Segmentation in 3-D LGE-MRI
1.0 1.0
0.9 0.9
Dice Coefficient
Dice Coefficient
0.8 0.8
0.7 0.7
0.6 0.6
1 Slice/Orientation 1 Slice/Orientation
0.5 3 Slice/Orientation 0.5 3 Slice/Orientation
5 Slice/Orientation 5 Slice/Orientation
0.4 0.4
1 2 3 4 5 6 7 8 9 1 2 3 4 5 6 7 8 9
Patient Number Patient Number
Figure 4.35: The 3-D interpolation evaluation results: (a) The A-HRBF result for
the endocard with an average Dice coefficient of 0.90, 0.94, and 0.95 for 1, 3, and 5
slices per orientation, respectively. (b) The A-HRBF result for the epicard with an
average Dice coefficient of 0.90, 0.94 and 0.95 for 1, 3, and 5 slices per orientation,
respectively.
at the left ventricular outflow tract). In this case, the normal vector computation
fails based on the previous method [Ijir 13], whereas using our method, the normal
vectors are oriented based on the contour extracted from the 2-D segmentation mask,
see Figure 4.37. This leads to a better segmentation result compared to the standard
HRBF method.
The runtime is an important factor for the employment in clinical practice. The
current implementation is single threaded only, without any optimization. The whole
segmentation pipeline needs less than 5 minutes, implemented with Python. The most
time-consuming step is the initialization of the left ventricle in the whole heart scan,
as a two-stage registration is applied.
The proposed filter based method achieves a Dice coefficient of 0.84 ± 0.04 for
the endocardium and 0.80 ± 0.06 for the epicardium. The learning based method
achieves a Dice coefficient of 0.84 ± 0.07 for the endocardium, and 0.85 ± 0.07 for
the epicardium. There is no big difference between the two proposed methods, as the
post-processing is similar for both methods.
The results for the base and the apical regions are worse compared to the mid-
cavity. For the base, this can be attributed to the fact that the transition from
the LV to the outflow tract is smooth. Therefore, the slice where the segmentation
algorithm stops is not defined by the volumetric data set, compared to regular 2-D
LGE-MRI. Furthermore, the clinical experts annotated the data in the axial, sagittal,
and coronal planes, so the transition of the outflow tract is even harder to delineate.
4.6 Discussion and Conclusion 105
Figure 4.36: The ground truth annotation in orange and the result of 3-D inter-
polation in red are shown for the endocard. The interpolation is obtained based on
one reference slices per orientation, where an overall DC of 0.92 is achieved. Each
row depicts a different orientation (sagittal, axial, and coronal) including the Dice
coefficient for each individual slice. It is expected that the closer to the reference
slice, the higher Dice coefficient is obtained.
In the apex, the endocardium gets very small. If there is a myocardial infarction
it is even harder to delineate between myocardium and blood pool, especially if the
annotation is not done in the short axis orientation. This is another reason for the
reduced DC in the apical region. Furthermore, for the data of the second cohort
no additional fat suppression is applied, therefore various enhancements in the apex
appeared. The last two columns of Figure 4.25 depict the apical region without scar,
however it can be seen that there the biggest differences occur compared to the gold
standard annotation.
Approaches as presented by Peters et al. [Pete 07] or Zhuang et al. [Zhua 10]
are not suitable for our case. First of all, the data sets differ. Furthermore, the
optimization of such models is difficult as there is no clear boundary between the
myocardial infarction and the blood pool. Therefore, no model based approach was
applied. To improve the results for the base, a cutoff parameter could be defined to
make an identical decision for the outflow tract. Furthermore, it can be seen that the
106 Left Ventricle Segmentation in 3-D LGE-MRI
Figure 4.37: Normal vector orientation for left ventricle segmentation with an am-
biguous boundary: (a) Control points in yellow and associated normal vectors in
blue based on intensity gradients for the HRBF [Ijir 13] method. (c) Control points
in yellow and associated normal vectors in blue based on the drawn contour in red
for our proposed A-HRBF method.
results for the epicardium segmentation of cohort two is worse compared to cohort
one. This can be explained by the insufficient fat suppression of the MRI data.
For the learning based segmentation, a graph cut based approach can be applied,
instead of dynamic programming to achieve a global cut for all the slices and not
only for each individual slice. This could lead to more consistent results throughout
the slices.
A comparison across different studies is difficult to perform, as data sets differ.
Also, most of the work reported in literature was performed using 2-D LGE-MRI.
Nevertheless, the proposed algorithm achieves similar or better results, in particular
when compared with the results by Albà et al. [Alba 14]. They reported a Dice
coefficient of 0.81 ± 0.05, solely using 2-D LGE-MRI. In the following the advantages
of the presented approach compared to the reported methods are identified.
First, Dikici et al. [Diki 04] and Wei et al. [Wei 11] only consider the 2-D slices
without taking the longitudinal axis into account. This can lead to inter-slice shifts
and to discontinuous 3-D shapes. In comparison, in this work the center, radius, and
shape of the previously found slices are considered, which allows for an inter-slice
smoothness.
4.6 Discussion and Conclusion 107
Second, Ciofolo et al. [Ciof 08] use a 3-D deformable mesh for the LV segmentation,
where the meshes are only attracted to features in the SA slices. As there are no
features considered in the other slices, this approach can lead to inter-slice shifts.
Wei et al. [Wei 13], who also deforms 3-D meshes, works in a 3-D framework and
considers the LA of the image. This approach adds more features for the deformation
and results in a smoother shape. Our method, while computational simple, uses the
pseudo SA slice and considers the propagation to the succeeding slices to deal with
potential misalignment and avoid segmentation errors, without the need to segment
any long axis or CINE images.
Although the gold standard annotation is obtained from a physician and a clini-
cal expert, it still involves inaccuracies regarding the manual delineation of the my-
ocardium. To some extent, this is due to a lack of good annotation tools. This issue is
also confirmed by the larger inter-observer variability. We have tried to compensate
this by an additional smoothing of the gold standard annotations. The effect of the
smoothing considering the Dice coefficient is rather minimal, as shown in Table 4.6.
The Dice coefficient changes only about 0.04 and the MSD is not notably affected
(less than 1 mm). However, the visual impression is improved. It has to be noted
that the smoothing is only in the pseudo short-axis slices and therefore, does not con-
sider the 3-D shape of the ventricle. Hence, this can lead to some stepping artifacts.
Furthermore, we could not compare our segmentation results to previous methods as
stated in Section 4.2, as there is none available for 3-D LGE-MRI.
For the precise scar quantification, a prerequisite is the accurate and reliable seg-
mentation of the left ventricle. However, the scar quantification heavily depends on
the technique applied [Kari 16, Rajc 14]. This is investigated in Chapter 5. Having
the segmentation of the scar, the myocardial infarction itself can be analyzed fur-
ther. For example, the infarct size and mass can be estimated, as these parameters
are good predictors for the success of a therapy. In addition to that, the type of
scarring is of interest. In clinical practice, three types of fibrosis are differentiated:
endocardial, epicardial, and transmural scar. Having the segmentation of the en-
docardium and epicardium, the transmurality of the scar can be classified precisely
[Reim 16, Reim 17a]. The different scar visualization methods are detailed in Chap-
ter 6.
The benefit of the filter basded and learning based method is the independence of
any user input. The algorithm requires only the 3-D LGE-MRI volume. This results
in a more robust segmentation. However, a graphical user interface is provided in
case the user is not satisfied with the segmentation result and the contours can be
edited manually.
For the semi-automatic left ventricle segmentation, our experiments showed that
one slice per orientation is sufficient to get a good segmentation result. Furthermore,
in order to achieve more accurate interpolation results, the user has to segment those
slices which have the maximum mismatch with the actual ground truth. In fact, for
3-D interpolation, the user selects those slices which are a good representation of the
complete shape. Hence, the actual result of the interpolation is even better than the
evaluation result shows.
In contrast to previous implicit methods for 3-D interpolation [Ijir 13], this method
can not only be used for high-contrast images, but also for images with high noise level
108 Left Ventricle Segmentation in 3-D LGE-MRI
or other confounding factors due to the independence of intensity information for the
3-D interpolation. The main advantage happens when there is an ambiguous bound-
ary which only an expert can recognize (e. g., between left ventricle and left atrium at
the left ventricular outflow tract). In this case, the normal vector computation fails
based on the previous method [Ijir 13], whereas using our method, the normal vectors
are oriented based on the contour extracted from the 2-D segmentation mask.
The purpose of this study was to provide an accurate and stable left ventricle
myocardial segmentation method for 3-D LGE-MRI sequences. Segmentation of the
LV endo- and epicardium has been studied in literature, but only a few methods
focused solely on LGE-MRI data. None of them considers 3-D LGE-MRI, to the best
of our knowledge. The presented work solely uses 3-D LGE-MRI for the segmen-
tation, unlike most related work which makes use of CINE MRI for the LGE-MRI
segmentation.
In the course of this work, two automatic and one semi-automatic segmentation
methods for the left ventricle endo- and epicardium have been presented that provides
accurate and consistent results for 3-D LGE-MRI. The automatic method achieved
an overall Dice coefficient of 0.84 for the endocard and 0.80 for the epicardium using a
simple filter based approach. Using the average surface distance, the endocard had a
mean error of 2.75 mm and the epicard had a mean error of 4.29 mm. A clear benefit
of the presented methods is the independence from an anatomical scan and from user
interaction.
Considering the semi-automatic segmentation, the method should be extended to
use arbitrary orientations of the 2-D slices, and not only axial, sagittal, and coronal
image slices for the 3-D interpolation. Having the possibility to annotate arbitrary
orientations, the 2-D annotations can be better adopted to the 3-D segmentation
problem. For the left ventricle for example, one would annotate the short axis ori-
entation, and the two long axis orientations to achieve excellent 3-D interpolation
results. Therefore, the 3-D interpolation with the 2-D gradient selector has to be
adopted. In addition, also the functionality of the 2-D smart brush can be improved,
as right now only the intensity distribution and the connectivity is considered. How-
ever, for medical images also the texture can play an important role. Incorporating
texture features for the classification of foreground and background could further
improve the 2-D segmentation result.
The benefit of the method is that the user can correct the segmentation result
easily by segmenting an additional slice with the maximum mismatch. Furthermore,
no prior knowledge is involved which leads to the ability to generate any arbitrary
segmentation of any 3-D data set, irrespectively of image modality, displayed organ,
or clinical application.
PART III
109
CHAPTER 5
In the previous two chapters, segmentation methods for the myocardium of the left
ventricle are presented for 2-D and 3-D LGE-MRI. In this chapter, the scar quantifi-
cation of LGE-MRI images is described. In Section 5.1, a short motivation is given.
Related literature is reviewed in Section 5.2. In Section 5.3, three different scar
quantification methods are outlined, namely, the x-standard deviation (x-SD), the
full-width-at-half-max (FWHM), and a texture based approach. The evaluation and
results of the different approaches are presented in Section 5.4. In the last section,
the results are discussed and a conclusion is drawn.
Parts of this chapter have been published previously in three conference publica-
tions [Kurz 15, Kurz 16a, Kurz 18b] and one journal article [Kurz 17f].
111
112 Scar Segmentation in LGE-MRI
5.1 Motivation
As previously mentioned, LGE-MRI is the clinical gold standard to non-invasively
visualize the viability of the myocardium. The contrast agent accumulates in the
damaged cells because of a delayed wash in and wash out in regions with ruptured
cell membrane, as there is an increased extracellular space. Furthermore, gadolinium
can diffuse in ruptured cell membranes [Dolt 13]. Both of these effects lead to an
increase in the gadolinium concentration and therefore is responsible for the hyper-
enhanced regions.
The quantification of the scar is very important for diagnosis, treatment plan-
ning, and guidance. Recent studies have shown that the knowledge about infarct
size, shape, and location can be very helpful during the ablation of ventricular tachy-
cardia [Estn 11, Andr 11]. In clinical routine, the scar is often segmented manually,
where hyper-enhanced tissue is selected. However, manual segmentation is prone to
inter and intra-observer variability and very time-consuming, especially for the 3-D
LGE-MRI acquisitions [Mirs 17]. Furthermore, there are also hyper-enhanced regions
at the location of the right ventricle insertion point due to the partial volume ef-
fect, especially at the left ventricular outflow tract and hyper-enhancements due to
epicardial and pericardial fat [Kari 16].
where the multiplication factor x for the standard deviation has to be defined by the
user, with x = {2, 3, ..., 6}.
The second common approach is the full-width-at-half-maximum approach, where
half of the maximum intensity within a user-specified hyper-enhanced region is se-
lected as the scar threshold [Amad 04]
However, both of these approaches require user input and are therefore still prone
to inter- and intra-observer variability. Hence, a fully automatic scar quantification
is desirable to achieve comparable results.
Tao et al. [Tao 10] combine intensity and spatial information for the scar segmen-
tation. First, an automatic thresholding is applied using Otsu’s method [Otsu 79],
where the intensities of the blood pool and the myocardium are considered. It is
expected that the histogram has two modes, one for the blood pool and the hyper-
enhanced tissue and one for the healthy myocardium. The blood pool is included in
the histogram as the Otsu’s threshold produces more stable results if there is only
a small myocardial infarction. Afterwards, the connectivity of the pixels is investi-
gated if less then three voxels are connected, these pixels are not further considered
as potential scar tissue. In addition, connected regions that are very thin in the long
axis direction are excluded.
Pop et al. [Pop 13] use a Gaussian mixture model in combination with an expec-
tation maximization to classify the intensity histogram of the myocardium into three
classes, healthy myocardium, border zone, and scar tissue.
Hennemuth et al. [Henn 08] combine a histogram analysis with a constrained wa-
tershed segmentation. Hence, some prior constraints are considered: i) The inten-
sity values of the LGE-MRI image are distributed according to the Rice distribu-
tion [Henn 08]. ii) The scar tissue is most likely sub-endocardial. iii) Relevant scar
tissue is compact and has a crescent-shaped area. iv) Dark regions surrounded by
scar tissue are no-reflow areas and are therefore included in the scar tissue.
Lu et al. [Lu 12] propose a graph cut based approach to segment the scar tissue in
LGE-MRI. The two terminal nodes are defined as healthy myocardium and scarred
myocardium, i. e., background and foreground of the image. The weights of the edges
are obtained from a Gaussian mixture model. For the delineation of the gray zone
and infarct zone, the FWHM algorithm is applied to the segmented infarct area.
Larrazo et al. [Larr 17] propose a classification based approach to segment the
myocardium. Therefore, texture features such as the run-length matrix, gray-level
co-occurrence matrix, and autoregressive model are extracted. In a feature selection
step, the most 17 discriminant features are selected and a support vector machine is
trained. To obtain the final segmentation result after the classification, morphological
opening is applied.
In general, a fully automatic scar segmentation is desirable, as the results are then
reproducible and not prone to inter- and intra-observer variability.
25000
20000
Number of pixels
15000
10000
5000
0
0 50 100 150 200 250 300
Intensities
Figure 5.1: Intensity histogram of the myocardium and blood pool. The blue line
indicates the Otsu’s threshold to separate the two modes. The green line indicates
the mean value of the lower mode, which is used to calculate the x-SD.
Number of pixels
20000
10000
0
0 100 200 300
Intensities
(a) 2-fold standard deviation
Number of pixels
20000
10000
0
0 100 200 300
Intensities
(b) 3-fold standard deviation
Number of pixels
20000
10000
0
0 100 200 300
Intensities
(c) 4-fold standard deviation
Number of pixels
20000
10000
0
0 100 200 300
Intensities
(d) 5-fold standard deviation
Number of pixels
20000
10000
0
0 100 200 300
Intensities
(e) 6-fold standard deviation
Figure 5.2: Segmentation result using the x-SD method for different values of x.
The first column, depicts the intensity histogram of the myocardium only and the
red dashed line indicates the x-SD threshold θxSD . The second column, depicts a
single slice, where the scar mask is overlaid in red. In the last column, the scar mask
is rendered in 3-D. It can be seen that with an increasing x value, the scar mask
decreases in size. The best segmentation result is achieved with x = 4.
116 Scar Segmentation in LGE-MRI
Number of pixels
20000
10000
0
0 100 200 300
Intensities
(a) FWHM threshold (b) 2-D slice (c) 3-D scar mask
Figure 5.3: Segmentation result using the FWHM method. (a) Depicts the intensity
histogram of the myocardium only and the orange dashed line indicates the FWHM
threshold θFWHM . (b) Illustrates a single slice, where the scar mask is overlaid in
orange. (c) The scar mask is rendered in 3-D.
line indicates the x-SD threshold θxSD . The second column illustrates a single slice,
where the scar mask is overlaid in red. In the last column, the scar mask is rendered
in 3-D. It can be seen that with an increasing x value, the scar mask decreases in
size.
Classification Phase
Scar
Image Pre- Feature Scar
Quan-
processing Extraction Classification
tification
normalization step of the myocardium is added, where the intensities within the
myocardium are also distributed in a range between 0 and 255.
Feature Extraction
In the next step, the features are extracted from the pre-processed image. For the
description of the scar within the myocardium, texture features are used. A fea-
ture vector f ∈ RD is created for every pixel within the myocardium, therefore, a
sliding-window approach is used. For each pixel also the surrounding neighborhood
is considered, with a pre-defined patch size of 7 × 7 pixels. For the feature extraction,
the segmentation-based fractal texture analysis (SFTA) algorithm is used [Cost 12].
The extraction algorithm is based on a set of binary images from which the fractal
dimensions of the regions’ borders are computed to describe the segmented texture
patterns. The fractal dimension provides an index of complexity comparing the de-
tails in a pattern by varying the scale where it is measured. The feature extraction
using the SFTA algorithm can be decomposed into two steps [Cost 12]. First, the
gray scale image is divided into a set of binary images. For the division, a technique
called two-threshold binary decomposition (TTBD) is applied. Second, for each of
the resulting binary images, the fractal dimension from the region boundaries is com-
puted. Furthermore, the mean gray level and the area of the remaining objects is
computed. These two steps are explained in more detail in the next two sections.
algorithm is to find the threshold that maximizes the intra-class variance. After the
first threshold is found, the Otsu algorithm is applied recursively to each sub-image,
until the the desired number of thresholds is reached. After the Nt threshold values
are defined, the input image I is decomposed into a set of binary images. Therefore,
pairs of threshold values θi and θi+1 are selected from the set of thresholds T , where
θi < θi+1 . In the next step, the pair of thresholds is applied to the input image I to
achieve a two-threshold segmentation
(
1 if θl < I(p) ≤ θu
I θ (p) = , (5.3)
0 otherwise
where θl and θu denote the lower and upper threshold, respectively. A set of binary
images is obtained by applying all possibles pairs of thresholds to the input image I.
These pairs consists of the continuous thresholds from T i and all pairs of thresholds
between θ and I max , where I max is the maximum intensity of the image I. Hence,
the number of binary images is 2Nt . The image shown in Figure 5.5 is decomposed
using the TTBD algorithm and visualized in Figure 5.6, where Nt = 8.
One important property of the TTBD is that the set of binary images obtained
from the algorithm is a super-set of all binary images that would be obtained by apply-
ing a one threshold segmentation using the thresholds computed with the multi-level
Otsu algorithm [Cost 12]. The aim of using threshold pairs for the image segmen-
tation is to segment structures that would not have been segmented using regular
thresholding. This is the case for gray values that lie in the middle ranges of the
input image.
SFTA Extraction Algorithm In the next step, the SFTA features are calculated
using the results of the two threshold binary decomposition. The SFTA feature
vector takes into account the size of the binary images, the mean gray value, and
the boundaries fractal dimension [Cost 12]. The fractal dimension is used to describe
the boundary complexity of the objects’ structure in the thresholded image I θ . The
boundaries of the thresholded image I θ are denoted as the border image ∆I. The
boundary image ∆I(x, y) has the value 1 if in the corresponding thresholded image
I θ at position (x, y) has the value 1 and at least one neighboring pixel has the value 0.
5.3 Scar Quantification 119
12 < I(p) ≤ I max 19 < I(p) ≤ I max 33 < I(p) ≤ I max 49 < I(p) ≤ I max
63 < I(p) ≤ I max 83 < I(p) ≤ I max 102 < I(p) ≤ I max 134 < I(p) ≤ I max
63 < I(p) ≤ 83 83 < I(p) ≤ 102 102 < I(p) ≤ 134 134 < I(p) ≤ 255
Figure 5.6: Decomposition of the image shown in Figure 5.5 using the TTBD
algorithm, where Nt = 8. The pairs of thresholds is given below each image.
120 Scar Segmentation in LGE-MRI
log Nι
4
−6 −4 −2 0
1
log
ι
Figure 5.7: The fractal dimension F is estimated by the slope of the regression line
of the estimated box counting values. In this case the fractal dimension is F = 1.46
Otherwise the boundary image ∆I(x, y) has the value 0. The aim of this procedure
is to have a one pixel thin boundary.
In the next step, the fractal dimension F ∈ R of ∆I(x, y) is computed using the
box counting algorithm, as binary images are used [Schr 92]. Therefore, the input
image ∆I is divided into a grid consisting of boxes of size ι × ι. In the next step, the
boxes are counted that contain at least one pixel of the object, resulting in Nι ∈ N,
which is depended on the box size ι ∈ N. The box size ι is varied till enough values
are obtained to estimate a curve defined by log Nι vs. log 1ι . Finally, the fractal
dimension F is computed by fitting a line to the curve to estimate, i. e., least square
fitting. In Figure 5.7 the estimated curve using the box counting values is visualized
with a fractal dimension of F = 1.46 for the first image of Figure 5.6. The fractal
dimension F corresponds to the slope of the line.
The mean gray value Ī and the area A ∈ R+ 0 of each binary image is extracted
without significantly increasing the computation time. Hence, the final feature vector
for each binary image I θ consists of three values, the fractal dimension F , the mean
gray value Ī, and the size of the binary image A. Thus, the final SFTA feature vector
consists of the number of binary images obtained from the TTBD times three. In the
case where Nt = 8, 16 binary images are obtained from the TTBD resulting in a final
SFTA feature vector of f ∈ R48 . The SFTA feature extraction pipeline is visualized
in Figure 5.8.
In addition to the SFTA features also the mean intensity I mean of the patch and
the local intensity I of the center pixel is extracted. Therefore, the final feature vector
used for the scar classification is of size f ∈ R50 .
For the classification of the scar tissue, a random forest classifier is used. The train-
ing of the random forest classifier is based on ground truth annotations from which
scar and non-scar samples are extracted. However, only pixels with the surround-
ing neighborhood completely inside or outside the infarcted myocardium are use for
training. Otherwise, information corresponding to both regions would be added to
the classifier.
5.4 Evaluation and Results 121
…
Figure 5.8: The SFTA feature extraction pipeline. First, the input image is decom-
posed in 2Nt binary images using the TTBD algorithm. Second, the SFTA features
are extracted, consisting of the fractal dimension F , the mean gray value Ī, and the
size of the binary image A for each binary image, respectively.
In this section, the evaluation of the three different scar quantification methods is
presented. First, the data used for the evaluation is described. In the second section,
the evaluation metrics are detailed. In the last section, the results are presented for
the three proposed methods.
122 Scar Segmentation in LGE-MRI
Figure 5.9: Example data set used for the evaluation of the scar quantification.
(a) Axial view of the data set, showing a small myocardial infarction (orange arrow).
(b) Coronal view and (c) sagittal view.
Table 5.2: Results of the scar quantification methods using the (a) Dice coefficient
and (b) the total volume error.
cross-validation is used, i. e., five data sets are used for testing and the remaining 25
are split into the training and validation data sets for the classifier. To optimize the
hyper-parameters, a grid-search is applied with the following parameters: number of
trees T ∈ {30, 40, 50, 60, 70, 80} and maximal tree depth D ∈ {5, 10, 15, 20, 25, 30}.
The inner loop of the nested cross-validation is set to a 5-fold cross-validation, i. e., five
data sets are used for the validation and the remaining 20 data sets are used for the
training of the classifier. The optimal hyper-parameters for the random forest are
T = 70 and D = 10.
5.4.3 Results
In this section, the results of the scar quantification are presented. The segmentation
using the learning based approach results in a Dice coefficient of 0.66 ± 0.17. The
best segmentation results in a DC of 0.85 and the worst in a DC of 0.11.
The Dice coefficient is also evaluated for the FWHM and the x-SD methods. The
results using the Dice coefficient for all the methods are summarized in Table 5.2 (a).
The worst result for the learning based scar quantification correlates with the result
for the FWHM method.
Furthermore, the DC of the different methods is visualized using a box plot, see
Figure 5.10.
In addition, the total volume error is evaluated for the learning based scar seg-
mentation. For the total volume error, a mean error of 0.04 ± 0.04 is achieved, with
a minimum error of 0.00 and a maximum error of 0.18. The TVE for all of the scar
quantification methods is summarized in Table 5.2 (b).
Furthermore, in Figure 5.11 (a) the individual Dice coefficients for the learning
based scar quantification are depicted in increasing order. The total volume error for
each data set is shown in Figure 5.11 (b) and sorted according to the Dice coefficient.
124 Scar Segmentation in LGE-MRI
1.0
0.8
Dice Coefficient
0.6
0.4
0.2
0.0
FWHM 2SD 3SD 4SD 5SD 6SD RF
Methods
Figure 5.10: Comparison of the Dice coefficient for all segmentation approaches,
where the blue line represents the mean Dice coefficient.
0.20
Total Volume Error
0.8
Dice Coefficient
0.15
0.6
0.10
0.4
0.2 0.05
0.0 0.00
1 5 10 15 20 25 30 1 5 10 15 20 25 30
Sequence Number Sequence Number
Figure 5.11: Segmentation results for learning based scar quantification. (a) In-
dividual Dice coefficients for learning based scar quantification in increasing order.
(b) Total volume error for each data set using the results of the learning based scar
quantification, sorted according to the Dice coefficient.
In addition, the correlation between the total volume error and the Dice coefficient
is investigated using a scatter plot. The result is depicted in Figure 5.12. The Pearson
correlation results in -0.16.
Moreover, the feature importance is investigated, as shown in Figure 5.13. The
most important feature is the SFTA feature. However, 48 features are compared to
two single features. Therefore, also the individual features of the SFTA are consid-
5.5 Discussion and Conclusion 125
1.0
0.8
0.4
0.2
0.0
0.0 0.2 0.4 0.6 0.8 1.0
Dice Coefficent
Figure 5.12: Scatter plot between DC and TVE of the learning based scar quan-
tification result, which results in a Pearson correlation of -0.16.
Importance
0.50
0.00
SFTA I Imean
Feature
Figure 5.13: Feature importance for SFTA features, mean intensity I mean , and the
local intensity I where the SFTA features are most important.
ered, see Figure 5.14. There it can be seen that the most important feature is the
mean intensity I mean followed by the mean gray value of different patches.
Besides the quantitative results, also the qualitative results of the proposed learn-
ing based method are evaluated. In Figure 5.15 the segmentation result is compared
to the gold standard annotation in 3-D and 2-D.
0.16
0.14
0.12
0.10
Importance
0.08
0.06
0.04
0.02
0.00
I
A7
A3
F2
F3
A2
A1
F4
A4
A8
A5
F5
F8
I¯2
A9
I1
I¯3
I¯1
I¯8
I¯7
I¯6
I¯4
I¯5
A16
F16
A6
F7
I¯9
F6
F1
A13
A10
F12
A15
A14
F15
F14
A12
F13
A11
F11
F10
I¯10
I¯12
I¯15
I¯11
I¯14
I¯13
Imean
¯
Feature
Figure 5.14: Detailed feature importance for all features in increasing order, where
the mean intensity I mean followed by the mean gray value of different patches are the
most important features.
total volume error. Normally you would expect a low TVE for a high Dice coefficient.
However, for data set 4, there is a low Dice coefficient but also a low total volume
error. After analyzing this data set, it can be seen that the total amount of scar is
very small. Therefore, the total overlap is bad, as there is only a small amount of
scar tissue within the myocardium, which is the reason for a low DC. Thus, this leads
to a small total volume error, as the volume itself is similar and only a small portion
of the myocardium has myocardial scar.
The benefit of all presented methods is the independence of any user interaction.
The algorithms can be applied to 2-D and 3-D LGE-MRI sequences. The purpose
of this study was to provide automatic, accurate, and stable tools for scar quantifi-
cation. The proposed learning based method achieves better results when compared
to state-of-the-art methods such as the x-SD method. In the course of this work, an
automatic scar quantification approach based on texture features has been presented
that provides accurate and consistent results for LGE-MRI sequences. Our method
achieves an overall Dice coefficient of 0.64 and a total volume error of 0.04. A clear
benefit of the presented methods is the independence from user interaction.
5.5 Discussion and Conclusion 127
Figure 5.15: Qualitative segmentation results for learning based scar quantification
compared to the gold standard annotation. (a) 3-D scar segmentation result for
the learning based scar segmentation. (b) 3-D gold standard annotation of the scar
tissue. (c) 2-D scar segmentation result for an individual slice. (d) 2-D gold standard
annotation for an individual slice.
128 Scar Segmentation in LGE-MRI
CHAPTER 6
Scar Visualization
6.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
6.2 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
6.3 Scar Layer Visualization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
6.4 Evaluation and Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
6.5 Discussion and Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
In the previous chapters, the segmentation of the left ventricle’s myocardium and
the scar quantification is detailed for 2-D and 3-D LGE-MRI. In this chapter, ad-
vanced scar visualization methods are described for both 2-D and 3-D, as an intuitive
visualization of the scar is helpful for planning and guiding the procedure.
In the first section, a motivation is given. In Section 6.2, an overview of related
work is provided. In Section 6.3, the two different methods for visualizing the scar
in 2-D and 3-D are outlined. The evaluation and the results for both methods are
presented in Section 6.4. In the last section, the results are discussed and a conclusion
is given.
Parts of this chapter previously appeared in two conference publications [Reim 17a,
Kurz 17g].
129
130 Scar Visualization
Figure 6.1: Two different scar visualization methods. (a) The scar is visualized as
a purple overlay on the MRI slice. (b) The scar is shown as a 3-D surface mesh in
purple and the endocardial mesh is shown in red.
6.1 Motivation
Heart failure affected in 2014 about 26 million people [Poni 14]. Cardiac resyn-
chronization therapy is one of the most successful therapies to treat patients with
advanced drug-refractory heart failure, systolic dysfunction, and ventricular dyssyn-
chrony [Moun 17]. However, the problem with CRT is that 30 % to 50 % of the
patients do not respond clinically to this therapy [Daub 12]. One of the reasons for
non-response is considered to be the suboptimal placement of the left ventricular
pacing lead. Pacing on areas of myocardial infarction has less or no effect, as the scar
tissue is not electrical conductive. Therefore, the localization and quantification of
scar tissue in the LV is crucial to increase the success rate of CRT [Reim 17a].
One important information is the scar burden and the scar transmurality. The
scar burden indicates the percentage of scar in the myocardium. If there is a high
percentage of scar burden, it is important to know if the scar is transmural, i. e., from
the endocardial wall to the epicardial wall. If the scar is only endocardial, it might be
still possible to place a lead in that segment. Hence, an intuitive scar visualization is
important to distinguish easily between endocardial, midcardial, and epicardial scar.
Figure 6.2: 2-D scar visualization. (a) The 3-D scar distribution in light gray is
projected on the BEP. (b) The scar burden is color coded depending on the percentage
of each segment. (c) The scar transmurality is color coded according to the thickness
of the scar in each segment.
distance from the apex along the principal axis of the LV as a proportion of the
height of the segmented area. The angle is defined by the right ventricle direction in
a plane which is perpendicular to the principal axis [Moun 17]. The problem with the
scar distribution is however, that it does not represent the scar location or thickness
within one segment of the BEP. Figure 6.2 (a) depicts the scar distribution in light
gray, where the 3-D scar mesh is projected on the segments of the BEP.
Hence, additional analysis has to be done to get the information of the scar burden
and the transmurality in each of the segments. The scar burden is the ratio between
myocardial scar and the total myocardial volume in each segment [Moun 17]. To
have an intuitive representation, the segments of the BEP are color coded according
the scar burden percentage, as depicted in Figure 6.2 (b). Furthermore, the scar
transmurality can be computed as the extent of scar as percentage of the myocardial
wall thickness, where the median of the transmurality is presented. For the intuitive
representation, the BEP is color coded depending on the transmurality percentage as
visualized in Figure 6.2 (c).
The drawback of the 2-D representations is the non-anatomical visualization of
the LV as a BEP and the missing information about the scar’s location within the
myocardium, as the transmurality gives no information if there is endocardial or
epicardial scar. The scar can either touch the epicardium, the endocardium, or
can be located within the myocardium without touching the epicardium or endo-
cardium [Reim 16]. Furthermore, no 3-D guidance is possible.
However, if scar is only endocardial, it might me still possible to place the lead
at that segment. Hence, an interactive scar location and transmurality visualization
method is proposed, where the scar is divided into different layers [Reim 17a].
Figure 6.3: Overview of the scar layer visualization pipeline, which is divided into
five main steps. First, the left ventricle’s myocardium and the scar are segmented.
Second, the epicardium and endocardium are delineated. Third, the layers are com-
puted. Fourth, the scar layers are extracted. In the last step, the scar layer’s can be
visualized in 3-D as meshes or projected onto the BEP in 2-D.
delineated from the segmentation masks. In the third step, the layers are computed,
i. e., one layer for endocardial, midcardial, and epicardial scar. Fourth, the scar layers
are extracted depending on the previously computed layer. Finally, the scar layers can
be visualized in 2-D and 3-D [Reim 17a, Kurz 17g]. An overview of the segmentation
pipeline is given in Figure 6.3.
r [pixel] 38
0◦ 90◦ 180◦ 270◦ 360◦
ρ
Figure 6.4: Transformed scar mask in polar coordinates with computed layers (red
and orange) between the endocardium and epicardium.
Figure 6.5: Scar layer generation process illustrated for one slice. (a) Segmenta-
tion mask with detected endocardium (red) and epicardium (yellow). Within the
myocardium, the scar is shown in a lighter shade of gray and white. (b) Depicts
the calculated layers, where l = 3. (c) The scar layers are generated using logical
comparison. (d) The final scar layer masks, which are used for the surface mesh
generation and the projection on the BEP.
and epicardium are computed. Therefore, the distance between the endocardium and
epicardium is calculated and then divided into multiple layers. For l ∈ N layers, the
myocardium needs to be divided l − 1 times. For each angle ρ, l − 1 values within
the myocardium are computed. The result of the layer generation is visualized in
Figure 6.4 in polar coordinates.
After the delineation of the l layers in polar space, they are transformed back
to Cartesian coordinated see Figure 6.5 (b). In this work, the number of layers l is
set to three. Because for the left ventricular lead placement, it is helpful to have
an epicardial, mid-myocardial, and endocardial scar layer to decide for the best lead
location.
In the fourth step, the scar layers are extracted. Therefore, the previously defined
layers and the scar mask are compared using logical operations. For the three scar
layers, the myocardium is divided twice, see Figure 6.5 (b). The first defined line is
next to the endocardial contour and the second defined line is close to the epicardial
contour. The first filled layer B 1 ∈ RN ×N ×N is defined as the area within the first
subdivision layer. The second filled layer B 2 ∈ RN ×N ×N is defined as the area within
the second subdivision line. The third layer B 3 ∈ RN ×N ×N is the area within the
epicardium. Then the filled layers are compared using logical operations with the
scar mask Z ∈ RN ×N ×N and the three individual scar layer masks Z 1 , Z 2 , and Z 3
are obtained, where
Z 1 = B1 ∧ Z , (6.1)
134 Scar Visualization
Z 2 = B1 ∧ B2 ∧ Z , (6.2)
and
Z 3 = B1 ∧ B2 ∧ B3 ∧ Z . (6.3)
The result is depicted in Figure 6.5 (c) and (d).
Afterwards, the scar layers can be visualized in 2-D or 3-D which is detailed in
Section 6.3.2 and Section 6.3.3.
The 3-D patient coordinate point is obtained by multiplying the pixel coordinates
with the affine matrix
xx xr
yy yc
= D . (6.6)
zy zs
1 1
The final result is shown in Figure 6.6, where in (a) all three layers are visualized,
in (b) the mid-myocardial layer and the endocardial layer, and in (c) only the en-
docardial layer. The subdivision of the scar mesh into several scar layers enables an
6.3 Scar Layer Visualization 135
Figure 6.6: 3-D scar layer visualization. (a) The endocardial mesh is visualized in
dark red with the three scar layers, from endocardial to epicardial. (b) Endocardial
mesh with the endocardial and mid-myocardial scar layer. (c) Endocardial mesh with
only one scar layer, the endocardial scar layer. (d) Standard representation of the
scar as one mesh in purple.
interactive peeling of the scar in 3-D. It can be scrolled from epicardium to endo-
cardium and vice versa. With the help of the scar layer visualization, an intuitive
representation of the transmurality of the scar is possible. If the scar is fully trans-
mural, all scar layers add up. Figure 6.6 (d) depicts the scar as one single mesh.
The 3-D scar layers can be overlaid onto the fluoroscopic images, as depicted in
Figure 6.7. This visualization method can be used during the intervention. For the
overlay, the epicardial mesh of the LV is registered to the fluoroscopic image [Toth 18].
Then, the epicardium, the endocardium, the scar mesh, and the scar layers can
be visualized in different colors. The colors and opacity can be adapted manually.
Meshes, which the physician is not interested in, can be hidden. This supports the
136 Scar Visualization
Figure 6.7: Fluoroscopic image with overlaid endocardial mesh in red and the three
3-D scar layers.
physician during the intervention, as only the required and important information is
shown.
Having the covariance matrix Σ, the SVD is applied to the covariance matrix for the
purpose of performing principal component analysis,
Σ = U SU T , (6.9)
Figure 6.8: Projection of the left ventricle onto the 16 segments BEP. (a) Estimation
of the principal axis of the left ventricle. The red point marks the mitral valve plane.
(b) The endocardial mesh of the left ventricle is projected to the 16 segment BEP.
(c) 16 segment BEP where the mitral valve plane is marked in red. Furthermore, for
orientation purposes, the septum, anterior, inferior, and lateral orientation is labeled.
the largest variation among the left ventricle’s endocardium, i. e., the short axis ori-
entation. Furthermore, the lowest and highest vertices, i. e., the LV apex and base,
are computed by projecting all the vertices v to the principal axis. In addition, the
mitral valve is automatically detected and excluded from the segment calculation, as
depicted in Figure 6.8 (b), where the red dot marks the mitral valve plane. The prin-
cipal axis is divided into three sections: the basal, the midcardial, and apical section.
Based on this, the mesh vertices in the basal and midcardial sections are divided into
six sub-sections, based on anatomical landmarks. The anatomical landmarks used
are the insertion points connecting the right ventricular and left ventricular wall. The
apical section is divided into four sub-sections to match the segments of the BEP.
The 16 segment BEP is depicted in Figure 6.8 (c).
After the division of the endocardial mesh into the 16 segments, also the scar
mesh is divided into these 16 segments. However, if the whole 3-D mesh is projected
into the BEP, a distinction between endocardial or epicardial scar is not possible as
seen in Figure 6.9 (b). Also the additional information about the scar transmurality
cannot help in distinguishing the location of the scar within the myocardium as
shown in Figure 6.9 (a). Therefore, the previously generated scar layers are projected
onto the BEP. The scar layers can be projected on top of each other as visualized
in Figure 6.9 (c). Alternatively the individual layers can be shown separately, as
depicted in Figure 6.9 (d)-(f) for the endocardial, midcardial, and epicardial layer,
respectively.
Figure 6.9: 2-D scar layer visualization. (a) Scar transmurality shown as percent-
age for each segment and color coded accordingly. (b) The 3-D scar distribution is
projected on the 16 segment BEP. (c) Three scar layers projected onto the 16 seg-
ment BEP, where a differentiation of endocardial, midcardial, and epicardial scar is
possible. (d) Endocardial scar layer. (e) Midcardial scar layer. (f) Epicardial scar
layer.
of the data please refer to Section 3.4.1. For the evaluation, two tests are created. In
the first test, nine physicians are shown four cases. For each case, two visualization
methods are presented: the segmented LV overlaid with the 3-D scar mesh and the
segmented LV overlaid with the scar layer visualization, similar to Figure 6.6. They
are asked to decide which visualization method they would prefer. In 80.6 % cases,
the clinical experts prefer the scar layer visualization, in 16.7 % cases they prefer the
3-D scar mesh and in 2.8 % cases they do not have a preference.
In the second test, eight physicians are shown six 3-D scar meshes and six scar layer
meshes. For each visualization method, they should decide if the scar is epicardial or
endocardial. The results are shown in Table 6.2.
These two experiments show, that with the scar layer visualization, the clinicians
can easily choose an optimal lead placement location as they can decide whether the
scar is epicardial or endocardial.
Table 6.2: Evaluation with eight clinical experts and twelve scar meshes. They
should decide for each mesh if the scar is epicardial, endocardial, or they could not
determine.
ity of the scar within the myocardium is needed, as scar is electrically almost non-
conductive. The results show, that the clinicians could easier decide about the scar
location and transmurality using the layer visualization. The precise control over how
the scar transmurality is visualized, in 2-D and 3-D, allows the user to see the scar
location to the extent of transmurality. An interactive scrolling through scar layers
is realized, such that scar layers are added or removed from the visualization. The
endocardium and the scar meshes can be further overlaid onto fluoroscopic images.
The overlay of the meshes can be used to guide an intervention.
140 Scar Visualization
PART IV
141
CHAPTER 7
Outlook
7.1 Left Ventricle Segmentation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
7.2 Scar Segmentation and Visualization. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
In this chapter, new possibilities of future work are outlined, based on the presented
methods and previous discussions.
143
144 Outlook
Finally, the runtime of the segmentation pipeline for the 3-D LGE-MRI needs to
be improved. First, the initialization of the left ventricle can be improved by using
marginal space deep learning [Ghes 16]. Regarding the contour extraction, as of now
the apical contours are refined and afterwards the basal contours. This could be
performed in parallel.
Considering the semi-automatic segmentation approach based on HRBF, the
method should be extended to use arbitrary oriented slices for the 3-D interpolation.
For the left ventricle, the short axis and long axis orientations would be desirable to
achieve even better segmentation results. Therefore, the gradient selector has to be
adopted. In addition, also the smart brush can be improved by not only considering
intensity features.
Summary
Cardiovascular disease is the major cause of death worldwide. In more detail,
ischemic heart disease is the leading cause of death. This disease is in a close rela-
tionship to heart failure. In 2014, about 26 million people worldwide suffered from
heart failure. Patients that suffer from chronic heart failure can benefit from cardiac
resynchronization therapy (CRT). CRT is a successful treatment option, for patients
suffering from drug-refractory heart failure, have a wide QRS complex, and a reduced
left ventricular ejection fraction with less than 35 %. The bi-ventricular pacemaker
synchronizes the contraction pattern of the heart. The CRT device has three leads,
one placed in the right atrium, one in the right ventricle, and the last one through the
coronary sinus on the left ventricles myocardium. The problem, however, is that 30 %
to 40 % do not respond to this therapy. Therefore, for precise procedure planning,
information about the left ventricle’s anatomy and the scar distribution is needed.
Late gadolinium enhanced MRI is the gold standard for non-invasive assessment
of the tissue viability. Recently, this technology was extended to 3-D for a more
precise quantification of the left ventricles’ myocardium to the extent of myocardial
scarring. The principles of LGE-MRI and the difference between 2-D slice-selective
sequences and 3-D sequences is detailed in Chapter 2. In addition, a brief overview
is given on the image processing basics applied in this thesis, the pattern recognition
pipeline is outlined, and the random forest classifier is explained in more detail.
However, the challenge arises in the image analysis of these MRI sequences. Be-
cause, if the LV lead is placed on fibrosis, there will likely be no impact from the
pacing, as scar tissue is electrical non-conductive. The focus of this work was on left
ventricle segmentation, scar quantification, and scar visualization for late gadolin-
ium enhanced magnetic resonance imaging to provide information for the procedure
planning and guidance.
In Chapter 3, segmentation methods for 2-D LGE-MRI are presented. The seg-
mentation pipeline consists of four major steps. First, the left ventricle is initialized
in the mid-slice of the short axis stack using circular Hough transforms and a cir-
cularity measure. Second, a morphological active contours approach is used for the
rough estimation of the blood pool. Third, the endocardial contour can be refined
using either a filter based approach or a learning based approach. For the filter based
approach, the edge information and the scar probabilities are used to generate a cost
map and a minimal cost path search is applied to get the final result. For the learning
based approach, steerable features are extracted in a ray casting fashion and classified
145
146 Summary
using a random forest. The final result is obtained by combining the classification
result with the scar probability and a dynamic programming approach. Finally, the
epicardial boundary is extracted by using the information of the endocardial contour
and applying also either a filter based or learning based approach for the generation
of the cost map and a minimal cost path search for the final endocardial contour.
The presented method is evaluated on 100 clinical data sets and achieves similar or
better results when compared to literature. For the learning based segmentation, a
5-fold nested cross-validation is used. Hence, 20 data sets are used for the testing
of the classifier and the rest is used for the training and validation of the classifier.
The learning based segmentation also slightly outperforms the filter based segmen-
tation, especially considering the endocardial contour segmentation, where a mean
Dice coefficient of 0.82 for the endocardium and 0.81 for the epicardium is achieved.
In Chapter 4, segmentation methods for 3-D LGE-MRI are detailed. The auto-
matic segmentation is divided into five major steps. First, the left ventricle has to
be detected in the whole heart LGE-MRI scan. Therefore, a two-stage registration
based approach is used, consisting of a rigid registration followed by a non-rigid reg-
istration. Afterwards, the principle components of the left ventricle are estimated.
Knowing the principle axis of the left ventricle, the volume is rotated around this
axes to estimate the short axis view, which is commonly used for the segmentation
of the left ventricle. In the next step, the endocardial contour is refined using either
a learning based approach or a filter based approach. For the filter based approach,
the cost array for the boundary estimation is calculated from the edge information
and the scar probability. For the learning based approach, a trained random forest
classifier is used to estimate the boundary probability. This probability combined
with the scar probability results in the cost array. The final endocardial contour for
both approaches is obtained by a minimal cost path search in polar space. For the
succeeding slices, the center and the radius of the previous contour are considered
to achieve a smooth endocardial segmentation. Afterwards, the epicardial boundary
is estimated using the information from the endocardial boundary. For the filter
based approach, the closest edge to the endocardial boundary with increasing radius
is searched for. For the final result, the convex hull is estimated to achieve a smooth
looking contour. For the learning based approach, a trained random forest classifier
is applied to get the boundary probability for the possible boundary candidates. In
the next step, a dynamic programming approach in polar space is applied to obtain
the final result. The iterations stops, if the apex or left ventricular outflow tract
is reached. For the 3-D LGE-MRI, also the papillary muscles are segmented using
Otsu’s thresholding and a connectivity check. Finally, the contours are exported as
3-D surface meshes and used for further processing.
In addition to the fully automatic left ventricle segmentation approaches, a generic
semi-automatic segmentation method based on HRBF in combination with a smart
brush is introduced. First, the user segments some 2-D slices using the smart brush.
In the next step, the contours are extracted from the segmented masks and control
points are estimated. Based on the contour, the normal vectors of the control points
are calculated. For intersecting planes, close control points are merged and a 3-D
normal vector is estimated. Finally, our new formulation of HRBF is applied to
reconstruct the desired surface.
147
For the evaluation of the filter-based 3-D LGE-MRI segmentation, 30 clinical data
sets from two sites are used. Gold standard annotations of the endocardium and
epicardium are provided by two clinical experts for each of the data sets. The Dice
coefficient for the filter based segmentation resulted in an overlap of 0.83 ± 0.04 for
the endocardium and 0.80±0.05 for the epicardium. As the gold standard annotation
are performed in the coronal, axial, and sagittal view an additional smoothing step
is added in the short axis orientation for smoother looking results. The DC using
the smoothed gold standard annotations resulted in an overlap of 0.84 ± 0.04 for the
endocard and 0.80 ± 0.06 for the epicard. Furthermore, the left ventricle is divided
into three equal thirds, the base, the mid-cavity, and the apex. The Dice coefficient is
evaluated separately for these three individual parts. The best overlap is shown for the
mid-cavity with a mean DC of 0.88 ± 0.07. In addition, also the parameter variability
is evaluated to see the influence of all individual parameters. For the learning-based
segmentation, a nested cross-validation is used for the optimization of the hyper-
parameters of the random forest classifier. For the learning based segmentation only
the first cohort is considered. The DC resulted in an overlap of 0.84 ± 0.07 for the
endocard and 0.85 ± 0.07 for the epicard.
For the semi-automatic segmentation also the first cohort is used. First the smart
brush is evaluated. For each data set, 5 slices per orientation with 5 different positions
for each slice are extracted, leading to 75 patched for each data set. For most patients
an average Dice coefficient of 0.87 is achieved. However, the main problem with the
automatic evaluation of the smart brush is that normally the smart brush inherently
involves human interaction. Hence, it is expected that manual annotation using the
smart brush would even achieve better results. For the 3-D interpolation the same
data sets are used. For each data set 1, 3, and 5 slices per orientation are extracted,
which means to have a total number of 3, 9, and 15 segmented slices, respectively. The
semi-automatic segmentation using our A-HRBF interpolation achieved an average
Dice coefficient for the endocard of 0.90 ± 0.02, 0.94 ± 0.01, and 0.95 ± 0.01 for 1, 3,
and 5 slices per orientation, respectively. For the epicard, an average Dice coefficient
of 0.90 ± 0.02, 0.94 ± 0.02 and 0.95 ± 0.01 for 1, 3, and 5 slices per orientation was
achieved.
Having the segmented left ventricle the scar tissue within the myocardium is quan-
tified. The scar quantification is described in Chapter 5. First, state-of-the-art scar
segmentation methods are reviewed, such as the full-width-at-half-max algorithm or
the x-SD method. These methods are implemented in a fully automatic manner, to
eliminate inter- and intra observer variability errors. Furthermore, a texture based
scar segmentation algorithm is developed. Therefore, 50 features are extracted from
a scar patch using fractal analysis of the texture. The segmentation based fractal
texture analysis (SFTA) can be decomposed into two major steps. First, the gray
scale image is divided into a set of binary images using two-threshold binary decom-
position. Second, for each binary image, the fractal dimension, the mean gray level,
and the size is computed. For the final feature vector in addition to the SFTA fea-
tures, the mean intensity value of the extracted patch and the center intensity value
is added to the feature vector. For the training of the random forest classifier, only
patches which are completely inside or outside of the scar tissue are used. The RF
classifier assigns each pixel a probability of corresponding to scar or healthy tissue.
148 Summary
In the next step, the probability image is binarized, small components are removed,
and morphological closing is applied. The proposed scar quantification method is
evaluated on 30 clinical LGE-MRI data sets. A 10-fold nested cross-validation is
used for the evaluation of the texture based scar quantification. The DC results in
0.63 ± 0.17 for the texture based approach. In addition, the total volume error is
evaluated, which results in 0.04 ± 0.04. The results of the texture classification is also
compared to the FWHM and the x-SD. It can be seen that the texture based scar
quantification outperforms the x-SD method. However, the results cannot directly
be compared to the FWHM method as the gold standard annotations are based on
this method.
In Chapter 6, the scar visualization is described. As for the CRT procedure plan-
ning, the scar transmurality plays an important role, a new scar layer visualization is
introduced. The scar layer workflow consists of five major steps. A pre-requisite, is
the prior segmentation of the myocardium and scar tissue. Afterwards, the anatomy
is delineated. In the next step, the layers between the endocardial and epicardial
boundary are computed. Having the separate layers, the scar is extracted for each
layer. Finally, the scar is visualized either in 3-D as surface mesh or in 2-D projected
onto the 16 segment bull’s eye plot of the American Heart Association. For the 3-D
visualization, the DICOM to patient transformation matrix has to estimated. All
information needed for the calculation of the matrix is in the DICOM header. For
the 2-D layer visualization, the principle axis of the left ventricle has to be estimated.
The principal axis is divided into three sections: the basal, the mid-cardial, and apical
section. Based on this, the layers in the basal and mid-cardial sections are divided
into six sub-sections, based on anatomical landmarks. The apical section is divided
into four sub-sections to match the segments of the BEP. The anatomical landmarks
used are the insertion points connecting the right ventricular and left ventricular wall.
With the scar layer visualization it is possible to distinguish between endocardial and
epicardial scar easily. Furthermore, also the transmurality of the scar can be inves-
tigated, and therefore, suitable pacing sites for the left ventricular lead of the CRT
device can be found.
In Chapter 7, ideas on future work are summarized to improve and further inves-
tigate the methods presented in this thesis. First, further improvement suggestions
for the left ventricle segmentation are given, such as the usage of a model, the detec-
tion of the left ventricular outflow tract, or the application of deep learning methods.
Afterwards, new ideas for the scar segmentation and visualization are proposed. In
addition, the next steps for the guidance of the CRT procedure are outlined, as the
segmented myocardium and scar has to be overlaid onto the fluoroscopic images.
Appendix
Contributions to Published Papers
In the course of this thesis, contributions were published in the form of conference
proceedings and journal articles. Being the first author, I was responsible for the
development of the methods, the implementation, the evaluation of the proposed
approaches, and the writing of the manuscripts. Thus I declare that the presented
work is my own, while gratefully acknowledging that I received valuable advice along
the way. This refers to the following publications which are part of this thesis:
149
150 Appendix
Furthermore, on two publications being the second author I was responsible for
the supervision of the students. I provided major support in the development of
the methods, the implementation, and evaluation of the proposed approaches. In
addition, I extensively assisted in the writing of the manuscripts including several
iterations of proof-reading. This refers to the following publications which are part
of this thesis:
CP Control Point
CV Cross-Validation
DC Dice Coefficient
ECG Electrocardiogram
Gd Gadolinium
HF Heart Failure
IR Inversion Recovery
LV Left Ventricle
153
154 List of Abbreviations
MI Myocardial Infarction
PD Proton Density
RF Random Forest
TE Echo Time
TI Inversion Time
TR Repetition Time
155
156 List of Algorithms
List of Figures
1.1 Causes of heart failure . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2 Normal and wide ECG signal . . . . . . . . . . . . . . . . . . . . . . 5
1.3 Illustration of a CRT device implant. . . . . . . . . . . . . . . . . . . 6
1.4 X-ray image acquired after a CRT implant. . . . . . . . . . . . . . . . 7
1.5 Graphical overview of thesis structure . . . . . . . . . . . . . . . . . . 12
157
158 List of Figures
General Symbols
c∈C Class label
r ∈ R+
0 Radius
x∈R x-Coordinate
y∈R y-Coordinate
z∈R z-Coordinate
N ∈N Node
T ∈N Size of forest
H:D→R Entropy
HS : D → R Shannon entropy
HG : D → R Gini impurity
p ∈ [0, 1] Probability
p ∈ RD D-dimensional point
ψ ∈ Rd Geometric primitive
161
162 Left Ventricle Segmentation
τ ∈ Rd Threshold inequality
dx ∈ R Tangent in x-direction
dy ∈ R Tangent in y-direction
e Edge function
gx ∈ R Gradient in x-direction
gy ∈ R Gradient in y-direction
k ∈ R+
0 Length of contour
rp ∈ R+
0 Radius
A ∈ R+
0 Area
Left Ventricle Segmentation 163
F Energy functional
b ∈ R2 Point vector
d ∈ R2 Tangent vector
g ∈ R2 Gradient vector
k ∈ R2 Curve
n ∈ R2 Normal vector
s ∈ R2 Weight vector
vi ∈ R3 3-D vertex
w ∈ R2 Weight vector
I ∈ RN ×M Image slice
R ∈ RN ×N N − D rotation matrix
164 Left Ventricle Segmentation
T ∈ RN ×N N − D rotation matrix
β ∈ R3 Weighting factor
κ∈R Curvature
λ1 ∈ N Weighting factor
λ2 ∈ N Weighting factor
ψ Function
θκ ∈ R Curvature threshold
θO ∈ R Otsu’s threshold
V ∈ RV Set of vertices
A ∈ R+
0 Area of binary image
F ∈R Fractal dimension
Nι ∈ N Number of boxes
Nt ∈ N Number of thresholds
V ∈R Scar volume
ei ∈ R3 Face indices
k ∈ R3 Translation vector
B ∈ RN ×N ×N Scar layer
Z ∈ RN ×N ×N Scar mask
∆s ∈ R Slice thickness
F ∈ RN Set of faces
θi ∈ R Threshold
167
168 List of Tables
Bibliography
[Abbe 17] E. Abbena, S. Salamon, and A. Gray. Modern differential geometry of
curves and surfaces with Mathematica. CRC press, September 2017.
[Abra 03] W. Abraham and D. Hayes. “Cardiac Resynchronization Therapy for
Heart Failure”. Circulation, Vol. 108, No. 21, pp. 2596–2603, November
2003.
[Adam 94] R. Adams and L. Bischof. “Seeded Region Growing”. IEEE Trans-
actions on Pattern Analysis and Machine Intelligence, Vol. 16, No. 6,
pp. 641–647, June 1994.
[Alba 14] X. Alba, F. i Ventura, M. Rosa, K. Lekadir, C. Tobon-Gomez, C. Hoogen-
doorn, and A. F. Frangi. “Automatic Cardiac LV Segmentation in MRI
Using Modified Graph Cuts with Smoothness and Interslice Constraints”.
Magnetic Resonance in Medicine, Vol. 72, No. 6, pp. 1775–1784, Decem-
ber 2014.
[Amad 04] L. C. Amado, B. L. Gerber, S. N. Gupta, D. W. Rettmann, G. Szarf,
R. Schock, K. Nasir, D. L. Kraitchman, and J. A. Lima. “Accurate
and Objective Infarct Sizing by Contrast-Enhanced Magnetic Resonance
Imaging in a Canine Myocardial Infarction Model”. Journal of the Amer-
ican College of Cardiology, Vol. 44, No. 12, pp. 2383–2389, December
2004.
[Amre 16] M. Amrehn, J. Glasbrenner, S. Steidl, and A. Maier. “Comparative Eval-
uation of Interactive Segmentation Approaches”. In: Bildverarbeitung für
die Medizin 2016, pp. 68–73, Springer, March 2016.
[Amre 17] M. Amrehn, S. Gaube, M. Unberath, F. Schebesch, T. Horz, M. Stru-
mia, S. Steidl, M. Kowarschik, and A. Maier. “UI-Net: Interactive Arti-
ficial Neural Networks for Iterative Image Segmentation Based on a User
Model”. In: C. Rieder, F. Ritter, I. Hotz, and D. Merhof, Eds., EG
VCBM 2017, pp. 143–147, September 2017.
[Andr 11] D. Andreu, A. Berruezo, J. T. Ortiz-Pérez, E. Silva, L. Mont, R. Bor-
ràs, T. M. de Caralt, R. J. Perea, J. Fernández-Armenta, H. Zeljko,
and J. Brugada. “Integration of 3D Electroanatomic Maps and Magnetic
Resonance Scar Characterization Into the Navigation System to Guide
Ventricular Tachycardia Ablation”. Circulation: Arrhythmia and Elec-
trophysiology, Vol. 4, No. 5, pp. 674–683, October 2011.
[Ange 05] E. Angelini, Y. Jin, and A. Laine. “State of the Art of Level Set Methods
in Segmentation and Registration of Medical Imaging Modalities”. In:
Handbook of Biomedical Image Analysis, pp. 47–101, Springer, 2005.
[Auri 11] A. Auricchio and F. W. Prinzen. “Non-Responders to Cardiac Resyn-
chronization Therapy – The Magnitude of the Problem and the Issues”.
Circulation, Vol. 75, No. 3, pp. 521–527, February 2011.
169
170 Bibliography
[Born 95] P. Börnert and D. Jensen. “Coronary Artery Imaging at 0.5 T Using
Segmented 3D Echo Planar Imaging”. Magnetic Resonance in Medicine,
Vol. 34, No. 6, pp. 779–785, December 1995.
[Boyk 01] Y. Boykov, O. Veksler, and R. Zabih. “Fast Approximate Energy Mini-
mization via Graph Cuts”. IEEE Transactions on Pattern Analysis and
Machine Intelligence, Vol. 23, No. 11, pp. 1222–1239, November 2001.
[CIBC 15] CIBC. 2015. Seg3D: Volumetric Image Segmentation and Visualiza-
tion. Scientific Computing and Imaging Institute (SCI), Download from:
http://www.seg3d.org.
[Ciof 08] C. Ciofolo, M. Fradkin, B. Mory, G. Hautvast, and M. Breeuwer. “Au-
tomatic Myocardium Segmentation in Late-Enhancement MRI”. In:
Biomedical Imaging: From Nano to Macro, 2008. ISBI 2008. 5th IEEE
International Symposium on, pp. 225–228, IEEE, May 2008.
[Cord 11] L. Cordero-Grande, G. Vegas-Sánchez-Ferrero, P. Casaseca-de-la
Higuera, J. A. San-Román-Calvar, A. Revilla-Orodea, M. Martín-
Fernández, and C. Alberola-López. “Unsupervised 4D myocardium seg-
mentation with a Markov Random Field based deformable model”. Med-
ical Image Analysis, Vol. 15, No. 3, pp. 283–301, June 2011.
[Cost 12] A. F. Costa, G. Humpire-Mamani, and A. J. M. Traina. “An Efficient
Algorithm for Fractal Analysis of Textures”. In: Graphics, Patterns and
Images (SIBGRAPI), 2012 25th SIBGRAPI Conference on, pp. 39–46,
IEEE, August 2012.
[Crem 07] D. Cremers, M. Rousson, and R. Deriche. “A Review of Statistical Ap-
proaches to Level Set Segmentation: Integrating Color, Texture, Motion
and Shape”. International Journal of Computer Vision, Vol. 72, No. 2,
pp. 195–215, April 2007.
[Crim 12] A. Criminisi, J. Shotton, and E. Konukoglu. “Decision Forests: A Unified
Framework for Classification, Regression, Density Estimation, Manifold
Learning and Semi-Supervised Learning”. Foundations and Trends
R in
Computer Graphics and Vision, Vol. 7, No. 2–3, pp. 81–227, March 2012.
[Crim 13] A. Criminisi and J. Shotton. Decision Forests for Computer Vision and
Medical Image Analysis. Springer Science & Business Media, January
2013.
[Dang 12] H. Dang, Y. Otake, S. Schafer, J. Stayman, G. Kleinszig, and J. Siew-
erdsen. “Robust methods for automatic image-to-world registration in
cone-beam CT interventional guidance”. Medical Physics, Vol. 39, No. 10,
pp. 6484–6498, October 2012.
[Daub 12] J.-C. Daubert, L. Saxon, P. B. Adamson, A. Auricchio, R. D. Berger,
J. F. Beshai, O. Breithard, M. Brignole, J. Cleland, D. B. DeLurgio,
K. Dickstein, D. V. Exner, M. Gold, R. A. Grimm, D. L. Hayes, C. Israel,
C. Leclercq, C. Linde, J. Lindenfeld, B. Merkely, L. Mont, F. Murgatroyd,
F. Prinzen, S. F. Saba, J. S. Shinbane, J. Singh, A. S. Tang, P. E. Vardas,
B. L. Wilkoff, and J. L. Zamorano. “2012 EHRA/HRS expert consensus
statement on cardiac resynchronization therapy in heart failure: implant
and follow-up recommendations and management”. Heart Rhythm, Vol. 9,
No. 9, pp. 1524–1576, September 2012.
[Dick 08] K. Dickstein, A. Cohen-Solal, G. Filippatos, J. McMurray, P. Ponikowski,
P. A. Poole-Wilson, A. Strömberg, D. Veldhuisen, D. Atar, A. Hoes,
A. Keren, A. Mebazaa, M. Nieminen, S. G. Priori, and K. Swedberg.
“ESC Guidelines for the diagnosis and treatment of acute and chronic
heart failure 2008”. European Journal of Heart Failure, Vol. 10, No. 10,
pp. 933–989, October 2008.
[Dijk 59] E. Dijkstra. “A Note on Two Problems in Connexion with Graphs”.
Numerische Mathematik, Vol. 1, No. 1, pp. 269–271, June 1959.
Bibliography 173
[Heim 09] T. Heimann and H.-P. Meinzer. “Statistical shape models for 3D medical
image segmentation: A review”. Medical Image Analysis, Vol. 13, No. 4,
pp. 543–563, August 2009.
[Hend 03] A. Hendrix and J. Krempe. “Magnete, Spins und Resonanzen-Eine Ein-
führung in die Grundlagen der Magnetresonanztomographie”. Erlangen:
Siemens AG, 2003.
[Henn 08] A. Hennemuth, A. Seeger, O. Friman, S. Miller, B. Klumpp, S. Oeltze,
and H.-O. Peitgen. “A comprehensive approach to the analysis of contrast
enhanced cardiac MR images”. IEEE Transactions on Medical Imaging,
Vol. 27, No. 11, pp. 1592–1610, November 2008.
[Hu 13] H. Hu, H. Liu, Z. Gao, and L. Huang. “Hybrid segmentation of left ven-
tricle in cardiac MRI using gaussian-mixture model and region restricted
dynamic programming”. Magnetic Resonance Imaging, Vol. 31, No. 4,
pp. 575–584, May 2013.
[Hu 14] H. Hu, Z. Gao, L. Liu, H. Liu, J. Gao, S. Xu, W. Li, and L. Huang. “Au-
tomatic Segmentation of the Left Ventricle in Cardiac MRI Using Local
Binary Fitting Model and Dynamic Programming Techniques”. PloS one,
Vol. 9, No. 12, p. e114760, December 2014.
[Huan 11] S. Huang, J. Liu, L. C. Lee, S. K. Venkatesh, L. L. San Teo, C. Au,
and W. L. Nowinski. “An Image-Based Comprehensive Approach for
Automatic Segmentation of Left Ventricle from Cardiac Short Axis Cine
MR Images”. Journal of Digital Imaging, Vol. 24, No. 4, pp. 598–608,
August 2011.
[Hwan 14a] T. Hwang, E. Girard, T. Kurzendorfer, X. Zhu, and A. M. Cahill.
“First Experience with iGuide Navigational Software Application for
Bone Biopsies in Pediatric Interventional Radiology”. Pediatric Radi-
ology, Vol. 44, No. 1, May 2014.
[Hwan 14b] T. Hwang, T. Kurzendorfer, E. Girard, X. Zhu, and A. M. Cahill. “First
Experiene with iGuide Navigational Software Application for Bone Biop-
sies in Pediatric Interventional Radiology”. In: S. 2014, Ed., Journal of
Vascular and Interventional Radiology, March 2014.
[Ijir 13] T. Ijiri, S. Yoshizawa, Y. Sato, M. Ito, and H. Yokota. “Bilateral Her-
mite Radial Basis Functions for Contour-based Volume Segmentation”.
In: Computer Graphics Forum, pp. 123–132, Wiley Online Library, May
2013.
[Jian 14] K. Jiang and X. Yu. “Quantification of regional myocardial wall motion
by cardiovascular magnetic resonance”. Quantitative Imaging in Medicine
and Surgery, Vol. 4, No. 5, p. 345, September 2014.
[Joll 11] M.-P. Jolly, C. Guetter, X. Lu, H. Xue, and J. Guehring. “Automatic
Segmentation of the Myocardium in Cine MR Images Using Deformable
Registration”. In: STACOM, pp. 98–108, Springer, September 2011.
[Joll 86] I. Jolliffe. Principal Component Analysis. Wiley Online Library, 1986.
[Kais 14] M. Kaiser, M. John, T. Heimann, T. Neumuth, and G. Rose. “Improve-
ment of Manual 2D/3D Registration by Decoupling the Visual Influence
of the Six Degrees of Freedom”. In: Biomedical Imaging (ISBI), 2014
IEEE 11th International Symposium on, pp. 766–769, IEEE, April 2014.
176 Bibliography
[Kell 12] P. Kellman and A. Arai. “Cardiac Imaging Techniques for Physicians:
Late Enhancement”. Journal of Magnetic Resonance Imaging, Vol. 36,
No. 3, pp. 529–542, September 2012.
[Tao 14] Q. Tao, S. R. Piers, H. J. Lamb, and R. J. van der Geest. “Automated Left
Ventricle Segmentation in Late Gadolinium-Enhanced MRI for Objective
Myocardial Scar Assessment”. Journal of Magnetic Resonance Imaging,
November 2014.
[Thev 08] P. Thévenaz, M. Bierlaire, and M. Unser. “Halt on Sampling for Image
Registration Based on Mutual Information”. Sampling Theory in Signal
and Image Processing, Vol. 7, No. 2, March 2008.
[Toma 98] C. Tomasi and R. Manduchi. “Bilateral filtering for gray and color im-
ages”. In: Computer Vision, 1998. Sixth International Conference on,
pp. 839–846, January 1998.
[Turk 02] G. Turk and J. F. O’brien. “Modelling with Implicit Surfaces that Interpo-
late”. ACM Transactions on Graphics (TOG), Vol. 21, No. 4, pp. 855–873,
October 2002.
[Wei 11] D. Wei, Y. Sun, P. Chai, A. Low, and S. H. Ong. “Myocardial Seg-
mentation of Late Gadolinium Enhanced MR Images by Propagation of
Contours from Cine MR Images”. In: Medical Image Computing and
Computer-Assisted Intervention–MICCAI 2011, pp. 428–435, Springer,
September 2011.
[Wei 13] D. Wei, Y. Sun, S.-H. Ong, P. Chai, L. L. Teo, and A. F. Low. “Three-
dimensional segmentation of the left ventricle in late gadolinium en-
hanced MR images of chronic infarction combining long-and short-axis
information”. Medical Image Analysis, Vol. 17, No. 6, pp. 685–697, Au-
gust 2013.
[Weis 08] D. Weishaupt, V. Köchli, and B. Marincek. How does MRI work?: An In-
troduction to the Physics and Function of Magnetic Resonance Imaging.
Springer Science & Business Media, October 2008.
184 Bibliography