Regularized Image Reconstruction in

Parallel MRI with MATLAB 1st Edition

Joseph Suresh Paul
Visit to download the full and correct content document:
https://textbookfull.com/product/regularized-image-reconstruction-in-parallel-mri-with-
matlab-1st-edition-joseph-suresh-paul/
More products digital (pdf, epub, mobi) instant
download maybe you interests ...

Design Optimization using MATLAB and SOLIDWORKS 1st

Edition Suresh

https://textbookfull.com/product/design-optimization-using-
matlab-and-solidworks-1st-edition-suresh/

Digital image interpolation in MATLAB First Edition Kok

https://textbookfull.com/product/digital-image-interpolation-in-
matlab-first-edition-kok/

Audio And Speech Processing With MATLAB Paul R. Hill

https://textbookfull.com/product/audio-and-speech-processing-
with-matlab-paul-r-hill/

MATLAB Parallel Computing Toolbox User s Guide The

Mathworks

https://textbookfull.com/product/matlab-parallel-computing-
toolbox-user-s-guide-the-mathworks/
MATLAB Image Processing Toolbox User s Guide The
Mathworks

https://textbookfull.com/product/matlab-image-processing-toolbox-
user-s-guide-the-mathworks/

Management Of Banking And Financial Services 4th

Edition Padmalatha Suresh & Justin Paul

https://textbookfull.com/product/management-of-banking-and-
financial-services-4th-edition-padmalatha-suresh-justin-paul/

Medical Image Processing, Reconstruction and Analysis-

Concepts and Methods, Second Edition Jiri Jan (Author)

https://textbookfull.com/product/medical-image-processing-
reconstruction-and-analysis-concepts-and-methods-second-edition-
jiri-jan-author/

Image Processing and Pattern Recognition Based on

Parallel Shift Technology First Edition Bilan

https://textbookfull.com/product/image-processing-and-pattern-
recognition-based-on-parallel-shift-technology-first-edition-
bilan/

Fortran 2018 with Parallel Programming 1st Edition

Subrata Ray (Author)

https://textbookfull.com/product/fortran-2018-with-parallel-
programming-1st-edition-subrata-ray-author/
 Regularized Image
Reconstruction in Parallel
MRI with MATLAB®
 Regularized Image
Reconstruction in Parallel
MRI with MATLAB®

Joseph Suresh Paul and Raji Susan Mathew

Medical Image Computing and Signal Processing Laboratory
Indian Institute of Information Technology and Management-Kerala
Trivandrum, India
MATLAB® is a trademark of The MathWorks, Inc. and is used with permission. The MathWorks does not warrant the accu-
racy of the text or exercises in this book. This book’s use or discussion of MATLAB ® software or related products does not
constitute endorsement or sponsorship by The MathWorks of a particular pedagogical approach or particular use of the
MATLAB ® software.

CRC Press
Taylor & Francis Group
6000 Broken Sound Parkway NW, Suite 300
Boca Raton, FL 33487-2742

© 2020 by Taylor & Francis Group, LLC

CRC Press is an imprint of Taylor & Francis Group, an Informa business

No claim to original U.S. Government works

Printed on acid-free paper

International Standard Book Number-13: 978-0-815-36147-3 (Hardback)

This book contains information obtained from authentic and highly regarded sources. Reasonable efforts have been made
to publish reliable data and information, but the author and publisher cannot assume responsibility for the validity of all
materials or the consequences of their use. The authors and publishers have attempted to trace the copyright holders of all
material reproduced in this publication and apologize to copyright holders if permission to publish in this form has not been
obtained. If any copyright material has not been acknowledged please write and let us know so we may rectify in any future
reprint.

Except as permitted under U.S. Copyright Law, no part of this book may be reprinted, reproduced, transmitted, or utilized in
any form by any electronic, mechanical, or other means, now known or hereafter invented, including photocopying, micro-
filming, and recording, or in any information storage or retrieval system, without written permission from the publishers.

For permission to photocopy or use material electronically from this work, please access www.copyright.com (http://www.copyright.
com/) or contact the Copyright Clearance Center, Inc. (CCC), 222 Rosewood Drive, Danvers, MA 01923, 978-750-8400. CCC is a
not-for-profit organization that provides licenses and registration for a variety of users. For organizations that have been granted a
photocopy license by the CCC, a separate system of payment has been arranged.

Trademark Notice: Product or corporate names may be trademarks or registered trademarks, and are used only for
identification and explanation without intent to infringe.
Visit the Taylor & Francis Web site at
http://www.taylorandfrancis.com

and the CRC Press Web site at

http://www.crcpress.com
Contents

Preface ..............................................................................................................................................xi
Acknowledgements .................................................................................................................... xiii
Authors ...........................................................................................................................................xv

1. Parallel MR Image Reconstruction .....................................................................................1

1.1 Basics of MRI ....................................................................................................................1
1.1.1 Basic Elements of an MR System ......................................................................1
1.1.2 Static Magnetic Field B0 ......................................................................................1
1.1.3 RF Magnetic Field B1 ...........................................................................................2
1.1.4 RF Receiver...........................................................................................................2
1.1.5 Gradient Fields ....................................................................................................2
1.1.6 Slice Selection ......................................................................................................3
1.1.7 Generation of FID................................................................................................4
1.1.8 Imaging.................................................................................................................4
1.2 Nyquist Limit and Cartesian Sampling ....................................................................... 7
1.3 Pulse Sequencing and k-Space Filling ..........................................................................9
1.3.1 Cartesian Imaging ..............................................................................................9
1.3.2 k-Space Features ................................................................................................ 11
1.3.3 Non-Cartesian Imaging ................................................................................... 13
1.3.3.1 Data Acquisition and Pulse Sequencing ......................................... 13
1.3.3.2 Transformation from Non-Cartesian to Cartesian Data............... 14
1.4 Parallel MRI.................................................................................................................... 15
1.4.1 Coil Combination .............................................................................................. 16
1.5 MR Acceleration ............................................................................................................ 17
1.5.1 Acceleration Using Pulse Sequences .............................................................. 17
1.5.2 Acceleration Using Sampling Schemes .......................................................... 18
1.5.3 Under-Sampled Acquisition and Sampling Trajectories ............................. 19
1.5.4 Artifacts Associated with Different Sampling Trajectories ........................ 20
1.6 Parallel Imaging Reconstruction Algorithms ........................................................... 21
1.6.1 Image-Based Reconstruction Methods ..........................................................22
1.6.1.1 SENSE ...................................................................................................22
1.6.2 k-Space Based Reconstruction Methods ........................................................ 24
1.6.2.1 SMASH................................................................................................. 25
1.6.2.2 GRAPPA............................................................................................... 26
1.6.2.3 SPIRiT ................................................................................................... 29
1.6.2.4 Regularization in Auto-calibrating Methods .................................30
1.6.3 CS MRI ................................................................................................................ 31
1.6.3.1 CS-Based MR Image Reconstruction Model .................................. 32
1.6.3.2 Sparsity-Promoting Regularization ................................................. 33
1.6.4 CS Recovery Using Low-Rank Priors.............................................................34
1.6.4.1 Low-Rank CS-Based MR Image Reconstruction Model ...............34
References ............................................................................................................................... 36

v
vi Contents

2. Regularization Techniques for MR Image Reconstruction .........................................43

2.1 Regularization of Inverse Problems...........................................................................43
2.2 MR Image Reconstruction as an Inverse Problem ...................................................44
2.3 Well-Posed and Ill-Posed Problems ........................................................................... 45
2.3.1 Moore-Penrose Pseudo-Inverse ...................................................................... 45
2.3.2 Condition Number ........................................................................................... 45
2.3.3 Picard’s Condition............................................................................................. 46
2.4 Types of Regularization Approaches......................................................................... 47
2.4.1 Regularization by Reducing the Search Space ............................................. 47
2.4.2 Regularization by Penalization ...................................................................... 47
2.5 Regularization Approaches Using l2 Priors .............................................................. 48
2.5.1 Tikhonov Regularization................................................................................. 48
2.5.2 Conjugate Gradient Method............................................................................ 50
2.5.3 Other Krylov Sub-space Methods .................................................................. 52
2.5.3.1 Arnoldi Process .................................................................................. 52
2.5.3.2 Generalized Minimum Residual (GMRES) Method..................... 53
2.5.3.3 Conjugate Residual (CR) Algorithm................................................ 55
2.5.4 Landweber Method .......................................................................................... 55
2.6 Regularization Approaches Using l1 Priors .............................................................. 56
2.6.1 Solution to l1-Regularized Problems .............................................................. 58
2.6.1.1 Sub-gradient Methods....................................................................... 59
2.6.1.2 Constrained Log-Barrier Method .................................................... 60
2.6.1.3 Unconstrained Approximations ...................................................... 61
2.7 Linear Estimation in pMRI.......................................................................................... 67
2.7.1 Regularization in GRAPPA-Based pMRI ...................................................... 69
2.7.1.1 Tailored GRAPPA .............................................................................. 69
2.7.1.2 Discrepancy-Based Adaptive Regularization................................ 70
2.7.1.3 Penalized Coefficient Regularization ............................................. 71
2.7.1.4 Regularization in GRAPPA Using Virtual Coils ........................... 71
2.7.1.5 Sparsity-Promoting Calibration ....................................................... 72
2.7.1.6 KS-Based Calibration ........................................................................ 74
2.8 Regularization in Iterative Self-Consistent Parallel Imaging
Reconstruction (SPIRiT) ............................................................................................... 74
2.9 Regularization for Compressed Sensing MRI (CSMRI) ......................................... 75
Appendix................................................................................................................................. 79
References ............................................................................................................................... 79

3. Regularization Parameter Selection Methods in Parallel MR Image

Reconstruction ...................................................................................................................... 85
3.1 Regularization Parameter Selection........................................................................... 85
3.2 Parameter Selection Strategies for Tikhonov Regularization ................................ 87
3.2.1 Discrepancy Principle ...................................................................................... 88
3.2.2 Generalized Discrepancy Principle (GDP).................................................... 89
3.2.3 Unbiased Predictive Risk Estimator (UPRE) ................................................ 90
3.2.4 Stein’s Unbiased Risk Estimation (SURE) ..................................................... 90
3.2.5 Bayesian Approach ........................................................................................... 91
3.2.6 GCV .................................................................................................................... 92
3.2.7 Quasi-optimality Criterion.............................................................................. 93
3.2.8 L-Curve .............................................................................................................. 94
Contents vii

3.3 Parameter Selection Strategies for Truncated SVD (TSVD) ................................... 95

3.4 Parameter Selection Strategies for Non-quadratic Regularization ....................... 97
3.4.1 Parameter Selection for Wavelet Regularization .......................................... 97
3.4.1.1 VisuShrink ........................................................................................ 99
3.4.1.2 SUREShrink ...................................................................................... 99
3.4.1.3 NeighBlock ...................................................................................... 101
3.4.1.4 SUREblock....................................................................................... 101
3.4.1.5 False Discovery Rate ...................................................................... 102
3.4.1.6 Bayes Factor Thresholding............................................................ 103
3.4.1.7 BayesShrink .................................................................................... 104
3.4.1.8 Ogden’s Methods............................................................................ 105
3.4.1.9 Cross-validation ............................................................................. 106
3.4.1.10 Wavelet Thresholding.................................................................... 106
3.4.2 Methods for Parameter Selection in Total Variation (TV)
Regularization ................................................................................................. 106
3.4.2.1 PDE Approach ................................................................................ 107
3.4.2.2 Duality-Based Approaches ........................................................... 108
3.4.2.3 Prediction Methods ....................................................................... 112
References ............................................................................................................................. 114

4. Multi-filter Calibration for Auto-calibrating Parallel MRI ....................................... 119

4.1 Problems Associated with Single-Filter Calibration ............................................. 119
4.2 Effect of Noise in Generalized Autocalibrating Partially Parallel
Acquisitions (GRAPPA) Calibration ........................................................................ 119
4.3 Monte Carlo Method for Prior Assessment of the Efficacy of
Regularization ..................................................................................................... 120
4.4 Determination of Cross-over .................................................................................... 121
4.4.1 Perturbation of ACS Data for Determination of Cross-over ..................... 121
4.4.2 First Order Update of Singular Values......................................................... 122
4.4.3 Application of GDP ......................................................................................... 122
4.4.4 Determination of Cross-over ......................................................................... 123
4.5 Multi-filter Calibration Approaches......................................................................... 128
4.5.1 MONKEES........................................................................................................ 129
4.5.2 SV-GRAPPA ..................................................................................................... 132
4.5.3 Reconstruction Using FDR ............................................................................ 133
4.5.3.1 Implementation of FDR Reconstruction ..................................... 136
4.6 Effect of Noise Correlation ........................................................................................ 141
Appendix............................................................................................................................... 143
References ............................................................................................................................. 143

5. Parameter Adaptation for Wavelet Regularization in Parallel MRI ........................ 147

5.1 Image Representation Using Wavelet Basis ............................................................ 147
5.2 Structure of Wavelet Coefficients ............................................................................. 147
5.2.1 Statistics of Wavelet Coefficients................................................................... 148
5.3 CS Using Wavelet Transform Coefficients .............................................................. 150
5.3.1 Structured Sparsity Model ............................................................................ 151
5.3.1.1 Model-Based RIP ............................................................................ 151
5.3.1.2 Model-Based Signal Recovery ...................................................... 152
5.3.2 Wavelet Sparsity Model.................................................................................. 154
viii Contents

5.4 Influence of Threshold on Speed of Convergence and Need for

Iteration-Dependent Threshold Adaptation ........................................................... 155
5.4.1 Selection of Initial Threshold ........................................................................ 156
5.5 Parallelism to the Generalized Discrepancy Principle (GDP).............................. 156
5.6 Adaptive Thresholded Landweber .......................................................................... 159
5.6.1 Level-Dependent Adaptive Thresholding .................................................. 161
5.6.2 Numerical Simulation of Wavelet Adaptive Shrinkage CS
Reconstruction Problem ................................................................................ 161
5.6.3 Illustration Using Single-Channel MRI ...................................................... 163
5.6.4 Application to pMRI ....................................................................................... 165
5.6.4.1 Update Calculation Using Error Information from
Combined Image (Method I) .......................................................... 165
5.6.4.2 Update Calculation Using SoS of Channel-wise Errors
(Method II) ........................................................................................165
5.6.4.3 Update Calculation Using Covariance Matrix
(Method III) .......................................................................... 166
5.6.4.4 Illustration Using In Vivo Data ...................................................... 167
5.6.4.5 Illustration Using Synthetic Data .................................................. 172
Appendix............................................................................................................................... 174
References ............................................................................................................................. 176

6. Parameter Adaptation for Total Variation–Based Regularization

in Parallel MRI .................................................................................................................... 181
6.1 Total Variation–Based Image Recovery ................................................................... 181
6.2 Parameter Selection Using Continuation Strategies .............................................. 182
6.3 TV Iterative Shrinkage Based Reconstruction Model ........................................... 183
6.3.1 Derivative Shrinkage ..................................................................................... 185
6.3.2 Selection of Initial Threshold ........................................................................ 186
6.4 Adaptive Derivative Shrinkage ................................................................................ 187
6.5 Algorithmic Implementation for Parallel MRI (pMRI) ......................................... 189
Appendix............................................................................................................................... 198
References ............................................................................................................................. 209

7. Combination of Parallel Magnetic Resonance Imaging and Compressed

Sensing Using L1-SPIRiT ................................................................................................. 213
7.1 Combination of Parallel Magnetic Resonance Imaging and
Compressed Sensing .................................................................................................. 213
7.2 L1-SPIRiT ...................................................................................................................... 214
7.2.1 Reconstruction Steps for Non-Cartesian SPIRiT........................................ 216
7.3 Computational Complexity in L1-SPIRiT ................................................................ 217
7.4 Faster Non-Cartesian SPIRiT Using Augmented Lagrangian with
Variable Splitting ........................................................................................................ 218
7.4.1 Regularized Non-Cartesian SPIRiT Using Split Bregman Technique ........ 219
7.4.2 Iterative Non-Cartesian SPIRiT Using ADMM .......................................... 220
7.4.3 Fast Iterative Cartesian SPIRiT Using Variable Splitting ..........................222
7.5 Challenges in the Implementation of L1-SPIRiT ....................................................225
7.5.1 Effect of Incorrect Parameter Choice on Reconstruction Error ............... 226
Contents ix

7.6 Improved Calibration Framework for L1-SPIRiT ................................................... 227

7.6.1 Modification of Polynomial Mapping ......................................................... 227
7.6.2 Regularization Parameter Choice ................................................................ 228
7.7 Automatic Parameter Selection for L1-SPIRiT Using Monte Carlo SURE .......... 228
7.8 Continuation-Based Threshold Adaptation in L1-SPIRiT ..................................... 229
7.8.1 L1-SPIRiT Examples........................................................................................ 230
7.9 Sparsity and Low-Rank Enhanced SPIRiT (SLR-SPIRiT) ......................................234
References ............................................................................................................................. 236

8. Matrix Completion Methods ............................................................................................ 239

8.1 Introduction ................................................................................................................. 239
8.2 Matrix Completion Problem ...................................................................................... 239
8.3 Conditions Required for Accurate Recovery .......................................................... 240
8.3.1 Matrix Completion under Noisy Condition ............................................... 241
8.4 Algorithms for Matrix Completion .......................................................................... 241
8.4.1 SVT Algorithm ................................................................................................ 242
8.4.2 FPCA Algorithm ............................................................................................. 243
8.4.3 Projected Landweber (PLW) Method .......................................................... 243
8.4.4 Alternating Minimization Schemes............................................................. 244
8.4.4.1 Non-linear Alternating Least Squares Method........................... 245
8.4.4.2 ADMM with Nonnegative Factors ................................................ 246
8.4.4.3 ADMM for Matrix Completion without Factorization .............. 246
8.5 Methods for pMRI Acceleration Using Matrix Completion ................................. 248
8.5.1 Simultaneous Auto-calibration and k-Space Estimation........................... 249
8.5.2 Low-Rank Modeling of Local k-Space Neighborhoods ............................ 253
8.5.3 Annihilating Filter–Based Low-Rank Hankel Matrix Approach............ 255
8.6 Non-convex Approaches for Structured Matrix Completion Solution for
CS-MRI ......................................................................................................................... 259
8.6.1 Solution Using IRLS Algorithm .................................................................... 260
8.6.2 Solution Using Extension of Soft Thresholding ......................................... 261
8.7 Applications to Dynamic Imaging ........................................................................... 262
8.7.1 RPCA ................................................................................................................ 263
8.7.2 Solution Using ADMM .................................................................................. 263
References ............................................................................................................................. 265

MATLAB Codes ......................................................................................................................... 269

Index ............................................................................................................................................. 301
Preface

Image reconstruction in parallel magnetic resonance imaging (pMRI) is a mathematical

process that generates combined images using a limited number of data samples acquired
simultaneously from an array of receiver coils. Based on the nature of under-sampling
and trajectories employed for acquisition, several variants of reconstruction methods
have evolved since the advent of pMRI. Two major categories of reconstruction methods
are those requiring prior or self-calibration, and the iterative type methods which do
not require calibration data. Regularization methods form an integral part in both variants
of reconstruction approaches as they involve inversion of ill-posed matrices of large sizes.
This book provides a detailed discussion on the key issues and challenges relating to the
choice and usage of regularization methods applied to pMRI reconstruction algorithms.
The main motivation behind the writing of this book is to bring forward some of the
unique identifiable features of several types of regularization techniques applied to pMRI
reconstruction. The book summarizes key aspects required for judiciously choosing regu-
larization parameters required for artefact- and noise-free reconstruction, with relevance
to each of the specific reconstruction variants. A broad spectrum of pMRI reconstruction
algorithms are discussed along with case studies and MATLAB® codes. Case studies are
presented with experimentally acquired data from both Siemens and GE scanners using
different types of MR sequences.
This book targets two groups of readers. The first group includes researchers and stu-
dents involved in optimization and development of pMRI reconstruction methods. The lat-
ter group of readers includes the practitioners of medical imaging and MR technologists
working in MRI centers and industries who require knowledge of image reconstruction
for their projects. Readers are not presumed to have an advanced background in math-
ematics, but they are expected to have a basic understanding of linear algebra and prob-
ability theory. Readers of this book are also expected to have basic knowledge of signal
and image analysis.

MATLAB® is a registered trademark of The MathWorks, Inc. For product information,

please contact:

The MathWorks, Inc.

3 Apple Hill Drive
Natick, MA 01760-2098 USA
Tel: 508 647 7000
Fax: 508-647-7001
E-mail: [email protected]
Web: www.mathworks.com

xi
Acknowledgements

We are thankful to Professor Jose A Ramos, at Nova Southeastern University, for his
insightful comments on the contents of our book. We express our special gratitude to
Professor Frithjof Kruggel, Department of Biomedical Engineering, University of California,
Irvine; Professor Michael Braun, University of Technology, Sydney; and Professor Socrates
Dokos, Graduate School of Biomedical Engineering, The University of New South Wales,
Sydney, for their patience in proofreading parts of this manuscript. We also thank
Professor Mathews Jacob, University of Iowa, and Professor Albert M. Thomas, University
of California at Los Angeles, for longstanding support and collaboration. We are thankful
to Sruthi Raghoothaman, Medical Image Computing and Signal Processing Lab, Indian
Institute of Information Technology and Management–Kerala (IIITM-K), in India, for her
technical help in the theoretical development of continuation scheme. We also thank all
past and present members of the Medical Image Computing and Signal Processing Lab for
their assistance in writing this book. We thank and acknowledge the following scholars
for sharing codes for comparisons: K Dabov et al. for BM3D, E M Eksioglu for BM3D-MRI,
J Caballero et al. for DLTG, and M Lustig and J M Pauly for SPIRiT.

xiii
Authors

Joseph Suresh Paul is currently a professor at the Indian Institute of Information

Technology and Management–Kerala (IIITM-K), India. He obtained his PhD in Electrical
Engineering from the Indian Institute of Technology, Madras, India, in 2000. His research
is focused on magnetic resonance (MR) imaging from the perspective of accelerating
image acquisition, with the goal of enhancing clinically relevant features using filters inte-
grated into the reconstruction process. His other interests include mathematical applica-
tions to problems in MR image reconstruction, compressed sensing, and super-resolution
techniques for MRI. He has published a number of articles in peer-reviewed international
journals of high repute.

Raji Susan Mathew is currently pursuing her PhD in the area of magnetic resonance (MR)
image reconstruction. She received a bachelor degree in Electronics and Communication
Engineering from Mahatma Gandhi University, Kottayam, and a master’s degree in Signal
Processing from the Cochin University of Science and Technology, Kochi, in 2011 and 2013,
respectively. She is a recipient of the Maulana Azad National Fellowship (MANF) from the
University Grants Commission (UGC), India. Her research interests include regularization
techniques for MR image reconstruction and compressed sensing.

xv
1
Parallel MR Image Reconstruction

1.1 Basics of MRI

Magnetic resonance imaging (MRI) is a non-invasive medical imaging technique that
uses the magnetic property of materials for visualization of soft tissues. As opposed
to other imaging modalities, such as computed tomography (CT) or ultrasound, MRI
utilizes the electromagnetic properties of hydrogen nuclei present in fat and water mol-
ecules. Once hydrogen atoms are placed in a powerful external magnetic field B0 , the
nuclear spins are aligned with the direction of the magnetic field. As the number of
protons aligning with the field is slightly larger, there exists a net magnetization in the
direction of the main magnetic field [1]. The precession frequency of the spins is pro-
portional to the main magnetic field, that is, ω0 = γ B0 , where γ is a gyromagnetic ratio
constant. This precession frequency which is in the MHz range is called Larmor or reso-
nance frequency.

1.1.1 Basic Elements of an MR System

The MR system consists of four essential components, including:

1. A uniform, steady magnetic field B0 . In a typical clinical MR system, this magnetic

field strength is either 1.5 or 3.0 T. When an object to be imaged is placed in this
field, a macroscopic or bulk magnetization M0 is established.
2. A pulsed, high power, radio-frequency (RF) magnetic field B1. The magnetization
in some specific region or all of the volume is excited using this pulse, giving rise
to a detectable signal.
3. A switched, linear, magnetic field gradient G. This field is superimposed on the
uniform field B0 to provide the spatial differentiation required to form an image.
There are three orthogonal gradients ( Gx , Gy , Gz ) , which can combine to yield any
gradient direction in three-dimensional (3D) space.
4. RF receiver coil used to detect the MR signal from the sample. The received signal
is processed to produce the image.

1.1.2 Static Magnetic Field B0

During MR data acquisition, the patient lies in a strong static magnetic field. Typically, the
methods to generate this field can be broadly classified into three: fixed magnets, resistive
magnets in which current is passed through a traditional coil of wire, and superconducting

1
2 Regularized Image Reconstruction in Parallel MRI with MATLAB®

magnets. However, as fixed and resistive magnets are generally restricted to field strengths
below 0.4 Tesla (T), they cannot generate the higher field strengths necessary for high-
resolution imaging. Therefore, superconducting magnets are used in most high-resolution
imaging systems. The superconducting magnets are made from coils of superconduct-
ing wire, and they require the coils to be soaked in liquid helium to reduce their tempera-
ture to a value close to absolute zero.

1.1.3 RF Magnetic Field B1

The RF pulses required to resonate the hydrogen nuclei are generated using an MRI
transmitter coil. The range of frequencies in the transmit excitation pulse together
with the magnitude of the slice-selection gradient field decide the width of the imag-
ing slice. Typically, a transmit pulse produces an output signal with a relatively nar-
row bandwidth of ±1 kHz. The time-domain waveform that is required to produce
this narrow frequency band resembles a sinc function. This waveform is generated
digitally at baseband and then up-converted using a mixer to the appropriate centre
frequency. Very-high-speed, high-resolution digital-to-analog converters (DACs) can
be used for direct RF generation of transmit pulses up to 300 MHz Therefore, wave-
form generation and up-conversion over a broad band of frequencies can be accom-
plished in the digital domain.

1.1.4 RF Receiver
Due to the capability of RF receiver to process signals from multiple receiver coils, mod-
ern MRI systems use more than two receivers to process the signals. The bandwidth of
the received signal is dependent on the magnitude of the gradient field and is typically
less than 20 kHz. Modern receiver systems use high-input-bandwidth, high-resolution,
12–16-bit analog-to-digital converters (ADCs) with sample rates up to 100 MHz, which
eliminates the need for analog mixers in the receive chain by directly sampling the
signals.

1.1.5 Gradient Fields

The gradient field used in MRI is typically a small magnetic field applied along the longi-
tudinal axis linearly varying in space. In the presence of gradient fields, the total magnetic
field at a given location is given by:

B ( r ) = ( B0 + Gx x + Gy y + Gz z ) k = ( B0 + G ⋅ r ) k, (1.1)

Here, r = xi , yj , zk, denotes the spatial location of the spin, and the vector G = Gx Gy Gz 
the field gradient, with Gx , Gy and Gz representing the gradient strengths along x, y, and z
directions, respectively. As the overall main magnetic field varies in space, the spins inside
the volume have spatially dependent resonance frequencies, expressed as:

ω ( r ) = γ ( B0 + G ⋅ r ) (1.2)

The addition of a gradient field to the static B0 field enables the resultant total magnetic
field to vary linearly in space. The Larmor frequencies also vary with spatial location as
they are proportional to the field.
Parallel MR Image Reconstruction 3

1.1.6 Slice Selection

Slice selection is a technique of isolating a single plane in the object being imaged, by excit-
ing the spins in that plane only. This is achieved by applying an RF pulse, which affects
only a limited part of the nuclear magnetic resonance (NMR) spectrum, in the presence of
a linear field gradient along the direction in which the slice is to be selected. This can be
explained in the following three steps involving application of:

1. Gradient pulse: A gradient magnetic field is applied in the z-axis superimposed on

the background magnetic field. Since the frequency of precession depends on the
magnetic field, the nuclei will have different frequencies throughout the z-axis.
2. Slice-select pulse: The magnetization of the nuclei is flipped to the transverse
plane by the application of an RF pulse. The RF pulse frequency should be the
same as the Larmor frequency of the nuclei in order to flip the precession of the
nuclei. As the Larmor frequency of nuclei is different along the z-axis, a slice of
interest can be selected by altering the frequency of the RF pulse.
3. Reset pulse: Due to the variation of frequencies along the gradient, the nuclei
begin to precess out of phase. Therefore, the nuclei are to be reset before selecting
the next slice. This is achieved by temporarily reversing the gradient or applying
180° RF pulse to reverse the precessional frequencies.

These steps for slice selection are illustrated in Figure 1.1. There are certain factors which
affect the slice properties. These include RF pulse bandwidth, RF pulse frequency and
gradient strength. The RF pulse bandwidth is the range of frequencies within the pulse.
Large bandwidth results in the selection of a larger slice, and vice versa. Varying the RF
pulse frequency moves the selected slice up or down the z-axis. The size of the slice is
determined by altering the gradient strength.

FIGURE 1.1
Slice selection and selective excitation.
4 Regularized Image Reconstruction in Parallel MRI with MATLAB®

1.1.7 Generation of FID

Upon application of an RF pulse at the resonance frequency, the atoms absorb energy and
tilt the net magnetization onto the xy plane. Also, the spins associated with the individual
protons begin to rotate about the z-axis at the same phase. When the RF transmitter is
switched off, the protons return back to the equilibrium state, that is, from traverse plane
( xy ) to longitudinal plane ( z ) . This relaxation is characterized by the time constant called
spin-lattice relaxation time, denoted by T1. With Mz0 and Mz ( t ) denoting the equilibrium
magnetization and the longitudinal magnetization at time t, respectively, the recovery of
longitudinal magnetization is modeled as a function that grows exponentially with time
constant T1. This is expressed mathematically as:

(
Mz ( t ) = Mz 0 1 − e −t/T1 . ) (1.3)

Due to the spin-spin interactions and variations in B0 , the spins eventually lose their
transverse magnetization. The time that characterizes their return to the equilibrium state
is called spin-spin relaxation time, denoted by T2 [1]. With Mxy0 denoting the transverse
magnetization at equilibrium, the transverse magnetization Mxy is obtained as follows:

Mxy ( t ) = Mxy0 e −t/T2 . (1.4)

The changes in magnetization induce an alternating current in the receiver coils, and this
signal is referred to as free induction decay (FID). FID is a short-lived sinusoidal electro-
magnetic signal generated immediately following the 90° RF pulse. This signal is induced
in the receiver coil by the rotating component of the magnetization vector in the xy plane,
which crosses the coil loops perpendicularly. The FID is one among the four basic types of
NMR signals generated in different ways. The other types include the gradient echo (GRE),
induced using one RF pulse and gradient reversal; the spin echo (SE), using two RF pulses;
and the stimulated echo generated using the application of three or more RF pulses.

1.1.8 Imaging
In one-dimensional (1D) imaging, a frequency-encoding (FE) gradient pulse is applied
immediately after application of the 90° - RF pulse, as shown in Figure 1.2. The resulting
FID signal is simultaneously digitized along with application of a readout pulse. The read-
out is enabled synchronously with the FE gradient pulse. Application of this gradient
along the x-direction, modulates the precession frequency of spins as a linear function of
their positions along the x-axis. This causes the spin frequency to vary spatially:

ω ( x ) = γ ( xGx + B0 ) . (1.5)

FIGURE 1.2
Application of gradient in 1D imaging. Following the excitation with a 90° RF pulse, the FID signal is read off
in the presence of the FE gradient.
Parallel MR Image Reconstruction 5

The net signal generated from an ensemble of spins can be modeled as the summation
of their contributions from each location along the x-direction. In the case where the
spin density is modeled as a continuous function of space, the resultant signal can be
expressed as:
∞
s ( t ) = kM0 e − jγ B0t
∫
−∞
ρ ( x ) e −t/T2 ( x )e − jγ xGxt dx (1.6)

In the signal equation, kM0 can be removed by using appropriate normalization/scaling.

Also, with the assumption that T2  t , or by quadrature detection, the factor e − jγ B0t can be
removed. Thus, the new signal equation takes the following form:
∞
s (t ) =
∫
−∞
ρ ( x ) e − jγ xGxt dx (1.7)

Denoting k x = γ Gxt / 2π , and rewriting Equation (1.7) as a function of k x yields:

∞
R ( kx ) =
∫ −∞
ρ ( x ) e − j 2π kx x dx (1.8)

The integral in Equation (1.7) represents the Fourier transform of the spin density func-
tion. The acquired MR data can be thought of as being defined in the spatial frequency
domain (Fourier space). It follows that the spin density can be obtained by taking the
inverse Fourier transform of the signal as:

ρ ( x ) = F −1 {R ( kx )} . (1.9)

For two-dimensional (2D) image acquisition, the pulse sequence must involve two gradi-
ents. Suppose a 90° pulse excites spins within a transverse slice at a particular value of z.
In order to acquire a spatially differentiated signal from the slice, gradients in the x and y
directions are to be applied as shown in Figure 1.3. In contrast to 1D imaging, a Gy gradient
is additionally applied for a fixed time τ , before the start of the acquisition. Then the spins
along y are dephased by

θ = γ yGyτ . (1.10)

After switching off the gradient, the acquisition begins and the frequency differentiation
along the y axis is also turned off. Alternatively, the differentiation in the phase is acquired
over a short duration of the Gy pulse. Therefore, the Gy gradient is referred to as a phase-
encoding (PE) gradient.
The FID is sampled in the presence of Gx, after turning off the y gradient. The spatial
frequencies denoted by k x and k y can be defined as follows:

γ Gx t
kx =
2π
(1.11)
γ Gyτ
ky = .
2π
6 Regularized Image Reconstruction in Parallel MRI with MATLAB®

FIGURE 1.3
Application of gradients in 2D imaging. Following excitation with a 90° pulse, a phase-encoding gradient Gy
is applied in the y direction for a time τ . At the end of that time, the phase acquired by parts of the sample
along the y-axis will vary in proportion to the value of y. Then, the FID signal is read off in the presence of the
frequency-encoding gradient Gx .

For one FID signal, there is only one value of the vertical spatial frequency k y . The sequence
must be repeated several times to obtain other values of k y , with the values of Gy stepped
up or down with each onset of the FID signal generation. Once the required set of data is
obtained, the data can be arranged to form a 2D array. The signal equation in the 2D case
can then be expressed as:
∞ ∞
− j 2π ( k x x + k y y )
R ( kx , ky ) =
∫ ∫
−∞ −∞
ρ ( x, y ) e dxdy , (1.12)

where ρ varies in the xy-plane. The inverse 2D Fourier transform of the array in (1.12)
reconstructs the spin density in the image space as follows:

{
ρ ( x , y ) = F −1 R ( kx , k y ) . } (1.13)

The sampled signal is arranged along horizontal lines in the (k x k y )-plane, as shown in
Figure 1.4. The horizontal line with k y = 0 represents the FID signal sampled with Gy = 0.
For a gradient step ∆Gy , the separation between the sampled horizontal lines in the
(k x k y )-space is given by

γ∆Gyτ
∆k y = . (1.14)
2π

FIGURE 1.4
The trajectory in the k x k y plane resulting from the application of the Fourier technique of Figure 1.3. Each line
corresponds to an increment of Gy. The crosses denote the sample points.
Parallel MR Image Reconstruction 7

Until now, the discussion was limited to the acquisition in the first quadrant of the
(k x k y )-plane. Similarly, one could sample the fourth quadrant and complete the posi-
tive half-plane. This is achieved by allowing the Gy gradient to take on negative values.
The negative half-plane is acquired by allowing Gx to take negative values. Because each
step change in the gradient requires a new FID signal and the associated waiting time
TR, sampling all the quadrants becomes a lengthy process. In practice, when the gradient
pulses are rectangular, the spatial frequencies are obtained by integration over the gradi-
ent pulse duration. For time varying gradient pulses,

γ t
kx =
2π ∫ G (t ’) dt ’ ,
0
x (1.15)

and

τ
γ
ky =
2π ∫ G (τ ’) dτ ’.
0
y

1.2 Nyquist Limit and Cartesian Sampling

For better interpretation of Nyquist limits for k-space sampling, let us first recall the Nyquist
sampling theorem for a 1D time domain signal. Suppose the highest frequency component
for a 1D analog signal is fmax . According to the Nyquist theorem, the sampling frequency
must be at least 2 fmax , in order to perfectly recover the underlying signal. The multiplication
of the time domain signal with the sampling train is equivalent to a convolution in the fre-
quency domain, wherein the frequency spectrum of the 1D time domain signal is convoluted
with the Fourier transform of the impulse train. However, as the Fourier transform of the
impulse train is a scaled impulse train with twice the separation in time domain, the convo-
luted signal in the Fourier domain will have the frequency spectrum of the 1D signal repli-
cated at each of the impulse. Therefore, if the separation between samples in the time domain
is chosen as ts = 1 / 2 fmax , then the 1D signal can be recovered back from its corresponding
frequency spectrum by applying a low-pass filter to the frequency spectrum. In case of overs-
ampling also, the signal can be recovered as the separation between the impulses is larger.
However, if the sampling frequency fs is less than that of the Nyquist rate, then the frequency
spectra overlap and the signal cannot be recovered by using low-pass filtering.
Extension of this idea to MR acquisition needs a slightly different interpretation. This is
because sampling in MRI is performed in the k-space, which itself is a frequency domain.
Therefore, the bandwidth is interpreted as a quantity pertaining to the maximum spa-
tial information contained in the imaging object. In other words, this is simply the extent
of spatial support of the object in the imaging plane. Consider an elliptical object with
dimensions 2Wx and 2Wy along the x- and y-axes, respectively. The Nyquist sampling steps
in k-space would then be ∆k xN = 2π /2Wx and ∆k yN = 2π /2Wy , respectively. If the sampling
steps are reduced further below the Nyquist limits ∆k xN and ∆k yN , the repetition inter-
vals of the imaging area, referred to the field of view (FOV), would increase in relation
to the object dimensions. This is the case with oversampling when FOVx /2Wx > 1, and
FOVy /2Wy > 1. Figure 1.5 illustrates the repetitions of the imaging plane when the k-space
is sampled at the Nyquist rate and below. The leftmost panel shows the k-space, which
8 Regularized Image Reconstruction in Parallel MRI with MATLAB®

FIGURE 1.5
k-space sampling and Nyquist criterion. The k-space sampling interval ∆k x and maximum extent k xmax are related
to the FOVx and voxel size ∆x, respectively, of the reconstructed image in x-direction. Similarly, ∆k y and k ymax are
related to the FOVy and ∆y.

is band limited, and the corresponding full object in the image space. The band limits
in the x and y dimensions are FOVx and FOVy , respectively. The central panel shows the
sampling grid on the k-space, and the corresponding object space with FOVx = 2Wx and
FOVy = 2Wy . As in the 1D case, the object space is repeated periodically with a repetition
interval equal to the FOV. Here, the object images do not overlap, and hence the object of
interest can be recovered by application of a low-pass filter. The rightmost panel illustrates
the case in which FOVx < 2Wx and FOVy < 2Wy . In this case, the object images overlap, and
a perfect recovery of the object is not possible.
FOV depends on the strengths of magnetic gradients, their durations and the acquisi-
tion time. By acquiring data at uniform intervals ∆ts in time, the smallest FE step in the
k-space will be ∆k x = γ Gx ∆ts /2π . This in turn determines the size of FOV. FOV should be
set appropriately to cover the whole imaged object. Since the encoding function is peri-
odic, parts of the imaged object outside the FOV are misinterpreted by aliasing as being
inside the FOV. While designing an MRI acquisition, appropriate choices of the FOV and
the spatial resolution (i.e., voxel size) are to be determined. As the magnitude of the gra-
dient fields cannot be increased beyond a limit, the total acquisition time is proportional
to the number and length of the scan lines acquired, in the case of Cartesian imaging.
In this case, a larger FOV (or smaller voxels) in the PE direction requires more scan lines.
However, acquisition time is unaffected by the spacing between samples within a scan
line, and arbitrarily large FOV in the FE direction can be achieved for free. Therefore, the
longest dimension of the volume can be chosen as FE direction so that one can overs-
ample the k-space in that direction to avoid any possibility of aliasing.
Parallel MR Image Reconstruction 9

1.3 Pulse Sequencing and k-Space Filling

A pulse sequence refers to the repeated application of a set of RFs and gradient pulses
for which the time duration between the successive repetitions together with the gra-
dient pulse amplitude and shape will influence the resultant contrast and signal inten-
sities of the MR image. There are computer programs that control all the hardware
aspects of MRI acquisition process. Generally, the pulse sequences are defined using
the repetition time (TR), the echo time (TE), the inversion time (TI) (if using inversion
recovery) in milliseconds, and flip angle in case of a gradient echo sequence. TR is the
time between two consecutive RF excitation pulses. A long TR allows the magnetiza-
tion in tissues to relax and return to their original values. Alternatively, a short TR
results in the protons from some tissues not having fully relaxed to realign with the
main magnetic field before the next measurement. TE represents the time from the
centre of the RF pulse to the centre of the echo. Larger extent of dephasing resulting
from longer TE can lead to reduced signal intensities. In contrast with a shorter TE, the
spins will remain in phase for a longer time, thereby reducing the extent of dephasing.
In the inversion recovery sequence, a conventional SE sequence is preceded by a 180°
inverting pulse. Here the inversion time TI refers to the time between the 180° invert-
ing pulse and the 90° pulse. Flip angle is the angle by which the net magnetization
vector is rotated by an RF pulse into the transverse plane. Flip angle depends on the
particular pulse sequence utilized, and it is critical in determining the signal intensity
as well as the image contrast.
SE is the most common pulse sequence used in MR imaging based on the RF refocus-
ing of spins. This sequence uses 90° RF pulses to excite the magnetization and 180° RF
pulses to refocus the spins and generate signal echoes. A schematic representation of
pulse sequencing using SE and the corresponding k-space filling is shown in Figure 1.6.
The top panel shows application of a 90° RF excitation pulse. At this time point, there are
no gradient fields, and the k-space location is at the zero frequency point, as shown in the
right panel. In the next step, an FE gradient pulse with amplitude Gx is applied along
the x-direction. With this, the k-space point traverses toward the +k x direction. Panels
in the third row show the 180°RF pulse applied along the longitudinal axis that causes
rotation of spins in the (x, y)-plane about the x-axis. Correspondingly, the k-space location
shifts instantly from the +k x,max position to the −k x,max position on the extreme left side of
the k-space. To traverse the lower portion of k-space, a PE pulse with maximum negative
swing is applied, as shown in the bottom left panel. Together with this, an FE pulse of
amplitude Gx is applied for a duration TE. The refocusing RF pulse, together with the FE
gradient pulse, fills the echo samples along the lowest PE line. It is seen that the echo
amplitude is maximum as the k-space trajectory passes through the centre of the line (i.e.,
the k y -axis). Before the next repetition interval, the magnetizations relax and the k-space
location again starts from the centre. This process is repeated with application of Gy gra-
dients of different pulse amplitudes, sampled discretely at intervals ∆k y as dictated by the
Nyquist criterion.

1.3.1 Cartesian Imaging

In conventional MR data acquisition, MR imaging is performed by acquiring k-space
data along a Cartesian or rectilinear trajectory. Cartesian MRI refers to the acquisition of
a Cartesian grid of samples of the spatial Fourier transform of the bulk magnetization.
10 Regularized Image Reconstruction in Parallel MRI with MATLAB®

FIGURE 1.6
The steps in a typical SE pulse sequence. Corresponding to each step, the k-space trajectories and location are
shown in the right side.

Here data is sampled in a line-by-line fashion on a rectangular grid. The main benefit
of this type of sampling trajectory is that images can be directly reconstructed using fast
Fourier transform (FFT). The spacing and extent of the Cartesian grid have a direct relation
to the acquisition time, FOV and voxel size of the resulting image. In Cartesian acquisi-
tion, the number of PE steps can be reduced by an acceleration factor or reduction factor
R by increasing the distance between the k-space lines. If only one line is left between
two k-space lines, the MR data acquisition is said to be accelerated by a factor of R = 2.
The reduction of sampling along the PE direction of the k-space results in a reduced FOV
in that direction with associated fold-over artifacts.
Parallel MR Image Reconstruction 11

FIGURE 1.7
Cartesian k-space acquisition. (a): Linear phase-encoding order or standard MR acquisition, (b): low-high or
centric k-space order.

The position of the line being filled in the k y direction is determined by the PE gradient.
The amplitude of the PE gradient is incremented in steps such that the next adjacent line
in the k-space is filled successively starting from one edge of the k-space until the opposite
edge is reached. This is illustrated in Figure 1.7 and is referred to as a linear phase-encoding
order [2]. In some of the dynamic applications, including contrast enhanced angiography,
a different ordering of the PE steps is adopted to retain the contrast information at the
onset of acquisition. In that case, the PE gradient is incremented from zero, but with an
alternating sign, starting from the centre of the k-space in steps to acquire lines in extreme
edges. This type of k-space sampling in the PE direction is referred to as the centric phase
encode order or low-high k-space order.

1.3.2 k-Space Features

The central k-space region carries low spatial frequency information, and peripheral data
carries high-frequency information required to improve the spatial resolution. The nominal
spatial resolution of the image can be improved by extending the data collection farther
from the k-space origin. Since a large chunk of the image information is contained in the
low spatial frequencies, the addition of high spatial frequency information can only sharpen
the image without affecting the contrast or the basic shape features [3]. This is illustrated
in Figure 1.8 using k-space truncation with two square windows of different sizes. With a
smaller square window used for truncation, the reconstructed image is blurred due to loss of
high-frequency signals near the periphery of the k-space. However, with inclusion of higher
frequencies using a larger window, the resolution of the reconstructed image is improved.
A k-space line in 2D Fourier imaging refers to the samples of MR signal received at a par-
ticular PE level. Since the k-space position is encoded by the gradient amplitude, sampling
through the Fourier space is accomplished by changing the gradient shape and amplitude
over time. Every point in the raw data matrix carries partial information for the complete
image. The outer rows of the raw data matrix corresponding to the high spatial frequencies
provide information regarding the borders and contours of the image, whereas the inner rows
of the matrix with low spatial frequencies provide general contrast information of the image.
In real MR experiments, the k-space signal is acquired in the discretized form. Consequently,
the discrete Fourier transform (DFT) is used for image reconstruction from k-space.
12 Regularized Image Reconstruction in Parallel MRI with MATLAB®

FIGURE 1.8
k-space and image are Fourier transform (FT) pairs. Panels in the top row show a completely filled k-space and
the resulting image. The middle and bottom rows illustrate the k-space and image obtained after truncating the
k-spaces with large and small rectangular windows. The effects of truncation are clearly seen in the zoomed
versions of the images shown in the rightmost panels.

As inferred from Figure 1.8, with a greater net strength of the phase-encoding gradient,
the data is mapped to a farther position from the k-space centre. Based on the polarity
of the gradient pulse, the mapping can be either toward the upper or lower directions.
Thus, the duration of the PE gradient governs the location of data on the vertical axis of the
k-space. Alternatively, with a greater net strength of the FE gradient, the data samples tend
to be located farther from the k-space centre. The data is mapped to the right side of the
k-space if the gradient is positive, or in the left direction if the gradient is negative. Thus,
the strength of the FE gradient determines the location of data on the horizontal axis of
k-space. After the acquisition of the entire k-space, inverse Fourier transform can be used
to reconstruct the image. With uniform Cartesian acquisition, the k-space data is filled dur-
ing the scan so that one phase-encoding line is collected per TR. An extensive coverage of
the effect of discrete k-space sampling are provided in [1,4].
Scanning time increases with an increase in the number of PE lines. In the case of 3D
imaging, second and third dimensions are acquired using phase-encoding. Compared to
frequency-encoding, phase-encoding is a much slower process. As an example, with a
TR of 500 ms, collection of 256 PE lines require 128 seconds in comparison to 8 ms for
acquisition of 256 samples in one PE line [5]. This indicates that by reducing the number of
acquired PE lines, the data acquisition time can be significantly reduced. The image recon-
structed from full k-space can achieve the best quality as all signals have been sampled
to fill the entire k-space. Considering the scanning time required, k-space reconstruction
is carried out from under-sampled k-space, in which only a limited number of k-space
samples are acquired.
Another random document with
no related content on Scribd:
The Project Gutenberg eBook of
Henkivakuutusherroja
This ebook is for the use of anyone anywhere in the United States
and most other parts of the world at no cost and with almost no
restrictions whatsoever. You may copy it, give it away or re-use it
under the terms of the Project Gutenberg License included with this
ebook or online at www.gutenberg.org. If you are not located in the
United States, you will have to check the laws of the country where
you are located before using this eBook.

Title: Henkivakuutusherroja
Romaani

Author: Veikko Korhonen

Release date: October 21, 2023 [eBook #71925]

Language: Finnish

Original publication: Jyväskylä: K. J. Gummerus Oy, 1916

Credits: Juhani Kärkkäinen and Tapio Riikonen

*** START OF THE PROJECT GUTENBERG EBOOK

HENKIVAKUUTUSHERROJA ***
HENKIVAKUUTUSHERROJA

Romaani

Kirj.

VEIKKO KORHONEN

Jyväskylässä, K. J. Gummerus Osakeyhtiö, 1916.

I

Jeremias Kantele, entinen kansakoulun opettaja harppaa pitkin

askelin muuanna kevätiltana Aleksanterilla. Hänen kätensä ovat
työnnettyinä syvälle päällystakin taskuihin ja katse on suunnattu
johonkin epämääräiseen pisteeseen kaukana kadun päässä, joten
hän yhtämittaa törmää yhteen vastaan tulijoitten kanssa.

Jeremias Kantele on aamupäivällä saapunut pääkaupunkiin

henkisesti ja ruumiillisesti ylenmäärin kiusaantuneena. Hän on
toiminut täsmälleen vuoden henkivakuutusasiamiehenä
Korvenhonka yhtiössä ja siitä saanut juuri sen verran tuloja, että on
hengissä pysynyt. Mutta vakuutusagentilla täytyy olla moitteeton
puku, kiiltonahkakengät ja knalli, ja niihin ylellisyystavaroihin eivät
Kanteleen tulot ole riittäneet. Vatsakin on viimeaikoina pyrkinyt
painumaan selkärankaa vasten.

Nyt on hänen päähänsä iskenyt ajatus pyytää ylennystä virassaan,

ja siinä toivossa on hän haalinut kokoon rahat Helsingin matkaa
varten. Hikoiltuaan yhtiönsä toimitusjohtajan huoneessa puoli päivää
on hän saanutkin pyytämänsä ylennyksen, vieläpä paremmilla
palkkaeduilla, kuin olisi voinut toivoakaan. Uuden uutukainen
viisisatanen taskussaan on hän lähtenyt rakkaan yhtiönsä
toimitusjohtajan huoneesta ja kiertelee nyt suuren onnensa
huumeissa pitkin katuja ilman erikoista päämäärää.

— Ai perhana!

Taaskin sattui yhteentörmäys ja Kantele katsahti tällä kertaa oikein

onnettomuustoveriaan, joka piteli leukaansa sadatellen.

— No, eikö se ole Varsala? Päivää!

— Ka, sinä lempoko siinä. Kun kulkee kuin lehmä kadulla, alkoi
Varsala purkaa sisuaan.

— Mihin sinä menet ja mistä tulet? kysyi hän kohta hieman

lauhtuneemmin
Kanteleelta, joka oli pyörähtänyt kävelemään hänen rinnalleen.

— Tulen suoraan yhtiöni toimitusjohtajan luota, ilmoitti Kantele.

— Varmaankin saamasta nuhteita huonosta hankinnasta, arveli

Varsala.

— Etpäs nyt arvannutkaan.

Kantele löi näppiä ja jatkoi.

— Sain virkaylennyksen. Minä olen nyt »herra tarkastaja» niin kuin

sinäkin.

— Älä.

— No niin. Ja eikö liene palkkakin yhtä suuri kuin sinulla?

Viisisataa kuussa ja tantiemiä. Kantele löi uudelleen näppiä, tällä
kertaa aivan Varsalan nenän edessä jatkaen riemastuneena.
 — Loppui vihdoinkin se ainainen kituminen. Pitikin jo kulkea vatsa
hoikkana kuin kulkukoiralla, ja vaatteetkin alkoivat riippua riekaleina.
Vielä tänä aamuna olivat ajatukset synkeitä kuin yö, kun näin, miten
toiset tulivat tyytyväisinä ruokapaikoista, eikä itselläni ollut taskussa
kuin pari kuparikolikkaa. Mutta nyt… katsos…

Ja Kantele veti setelin povestaan ja näytti Varsalalle.

— Kas poikaa, kun ihan jo suuria rahoja näyttelee. Sinä sait

ennakkoa?

— Sain, Toimitusjohtaja antoi hienoja huomautuksia minun

ulkoasustani. Ja eihän tämä loistava olekaan, vaikka päällystakilla
itse asiassa on toinen omistaja. Huomautin hänelle yhtä hienosti,
että matti on hiipinyt taskuuni. Ja niinpä tuo kultainen mies nosti
minut silmänräpäyksessä liejusta, johon olin painumassa. Mutta
minun pitäisi saada ruokaa. Näes, tänään en ole vielä syönyt
muonaakaan. Mennäänpä Cataniin.

— Taitaa olla parasta mennä viivyttelemättä johonkin minun

ruokapaikoistani: Ala-Kämppiin tahi Oopperakellariin, virkkoi Varsala
merkitsevästi, antaen toisen näin tietää, että hän ei ainakaan nähnyt
nälkää silloin, kun sattui pyörähtämään pääkaupungissa.

He menivät Oopperakellariin ja saivat haltuunsa pöydän. Varsala

sytytti sikaarinsa ja virkkoi, vedettyään muutamia perinpohjaisia
savuja ja hetken niistä nautittuaan.

— Vai pääsi poika ylenemään. No, nythän sinä saat tuntea, miten
ihanaa on olla henkivakuutusherrana. Useimmiten toiset tekevät
työn, ja me saamme niittää sadon, ainakin mitä tantiemiin tulee.
Meidän on tosin pidettävä huoli kilpailusta ja siitä, että asiamiehet,
joille varsinainen työnteko kuuluu, eivät pääse laiskottelemaan,
mutta sehän tehtävä onkin kaikkein helpoimpia. Ja ajatteles, ethän
sinä enemmän kuin minäkään, tullessani Leimausyhtiön
palvelukseen ole suorittanut mitään tutkintoja tällä alalla, ainoastaan
ovelasti päättänyt muutamia tukkikauppoja maalaisten kanssa. Niistä
minut napattiin tähän toimeen. Huomasivat kai, että minussa on
liikeneroa ja että osaan olla ovela. Ja hyvin minä olen
menestynytkin, osaan herättää ihmisissä, varsinkin maalaisissa
luottamusta. Sen vaikuttaa ulkoasuni. Minä näes en komeile
vaatteilla. Olen sattunut usein toisten yhtiöitten samaan luokkaan
kuuluvien virkamiesten kanssa samaan aikaan hankinnalle ja tullut
huomaamaan, että kovin koreaa herrasmiestä katsellaan kansan
seassa hiukan sekavin tuntein. Minulla on siitä tuoreita esimerkkejä.
Viime viikolla yövyin erääseen taloon, jossa hieroin kymmenen
tuhannen vakuutusta, jota oli samana päivänä kärttänyt sinullekin
hyvin tunnettu Rientoyhtiön Keikaus. Kuulin miten isäntä arveli
emännälleen: — Tuo näyttää oikealta mieheltä! Se Keikaus oli niin
hieno mies ja isovatsainen. Taitaa syödä vakuutettaviensa varoja.
Mitäs, jos ottaisimmekin vakuutuksen Leimaus yhtiössä. Ja aamulla
se vakuutus päätettiin.

Kantele aukoili suutaan keskeyttääkseen puhetoverinsa

sanatulvan, sillä hänkin olisi tahtonut suuren riemunsa vallassa
sanoa jonkun sanan, vaan Varsala jatkoi.

— Kun sinäkin olet nyt ylennetty meidän joukkoomme, niin muista,

ettet kovin rupee koreilemaan, vaan pysyt alavana ja vertaisena
isäntien kanssa, ja saat nähdä, että sinua vetelee silloin kaikkein
parhaiten. Täytyyhän meidän itsemme kumminkin tehdä niitä
suurempia vakuutuksia ja muokata maata asiamiehille.
 — Sinä olet viisas mies, virkkoi Kantele ahmien ruoka-annostaan.

— Jaa. Sinäpä sen sanoit. Sitä niinä juuri olen. Ja jonkunverran

täytyy meissä kaikissa viisautta olla, jos mieli jotain saada aikaan.
Vaikkapa se viisaus meissä ilmeneekin niin sanottuna oveluutena,
niin joka tapauksessa sitä täytyy meissä olla vieläpä niin, että se on
huomattavissa jo päältä päin, sillä muutoin emme tähän asemaan
koskaan pääse kohoamaankaan. — Mutta, veli hyvä, tarvitaan tässä
toimessa muutakin ei vain neroutta, jota minun mielestäni kaivataan
kaikista vähimmin. Vakuutusherrana onnistuu minun mielestäni
parhaiten se, joka on jo syntyessään jonkun verran säikähtänyt. Vai
mitä?

— Jospa se lienee niin kuin sanot. Mutta mitä muuta sanoisit vielä
ihmisiä piinatessa tarvittavan?

— Tietysti häikäilemättömyyttä, oikein kouraan tuntuvaa. Täytyy

uskaltaa puhua asioista yhtä hyvin kerskalle valtiomiehelle kuin
vaikkapa villille ihmissyöjälle, eikä saa päästää uhria kynsistään,
ennen kuin on saanut hänet antautumaan.

— Et sinä ole hulluimpia vakuutusmiehiä, voit vielä tähtenä loistaa

joukossamme, virkkoi Varsala.

Virkaveljet laskeutuivat kadulle, jossa virtaili yleisöä soittoa

kuunnellen, Hetken kuluttua virkkoi Kantele Varsalalle.

— Tällä hetkellä minä voisin syleillä vaikka koko maailmaa. Eilen

illalla minulla oli taskussa kaksikymmentä penniä, ja nyt siellä on
seteli, jonka moista en ennen eläessäni ole omistanut Ja sitten se
suuri vapaus! Saa mennä mihin tahtoo, olla missä tahtoo. Saa syödä
joka päivä vatsansa täyteen ja vedellä makeita unia talojen
parhaimmissa kamareissa.

— Johan sinä innostut ennen iltaa, virkkoi Varsala. Jos tänään

juhlisimmekin sinun kunniaksesi, niin luulen ihmisen vasta puoliyön
tienoissa olevan valmiina syleilemään vaikka koko maailmaa. Mitä
taas tuohon elämän ylistykseesi tulee, niin onpa
henkivakuutusmiehellä vastuksiakin. Saa joskus varoa nahkaansa,
ja selkäsauna on toisinaan hyvin lähellä. Mutta yleensä on meidän
työmme helpon puoleista. Ja kukapa työn keveydestä voisi
nauttiakaan, jos ei saisi joskus olla kiipelissä. Tunnen tämän alan
jokseenkin täydellisesti. Olen kaksi vuotta rämpinyt Suomen
sydänmaita, ja kokenut siellä joskus karvastakin, vaikka tosin
makeampi puoli on useimmiten ollut minun osakseni suotu.

— Tästä puolin ne alkavat makean leivän päivät minullekin.

Varsala naurahti.

— Samoin kuin pahatkin päivät. Mutta kun selkäsaunan

saadessasi puret hammasta ja ajattelut, että tämä on
ylenmääräisestä tungettelevaisuudesta täysin ansaittua, niin hyvin
se menee.

Ystävykset pyörähtivät Prinsess’iin. Siellä oli Kanteleen

virkaylennys saapa juhlallisen vahvistuksen illan kuluessa.
II

Toukokuun päivä paistoi lämpimästi Kuivalan väen tupaan. Isäntä

kellotti sängyssään selällään koettaen saada unen päästä kiinni,
vaan siinä puuhassa häiritsivät häntä kärpäset, joita tänä keväänä oli
ilmestynyt ihmeteltävän aikaiseen. Ja ihmeellistä kyllä, ahdistelivat
ne pääasiassa vain isäntää ja talon vanhaa luppakorvakoiraa.
Ehkäpä syynä siihen oli se, että isäntä ja koira olivat talon asukkaista
laiskimmat.

Tuvasta oli emäntä poistunut taloustoimiinsa, piiat ja rengit

menneet pelloille. Isäntä oli päättänyt paneutua nukkumaan,
koskapa sattui hyvä tilaisuus, ja sitä varten ajanut lapset luudan
vartta heilutellen ulos tuvasta ja vielä pihastakin kartanon taakse.

Juuri kun isäntä oli pääsemässä uneen, astui tupaan Kantele,

Honkayhtiön uusi tarkastaja yhtä uuden eli vasta leivotun
asiamiehen
Alitalon Villen kanssa toivotellen hyvää päivää.

— Antakoon, murahti isäntä sängystä ja antoi silmäluomiensa

painua kiinni.

— Tuoko se on isäntä? kysyi Kantele Villeltä.

 — Se.

— Taitaa olla hyväluontoinen mies?

— No ei se ole pahimpiakaan. Onpahan semmoinen laiska mies,

totesi
Ville.

— Sitä on sitten sen mukaan lähestyttävä, neuvoi Kantele uutta

asiamiestä.

— Jaa, virkkoi Ville kunnioittavasti.

— No hyvinkö se nukuttaa, kysyi Kantele isännältä.

— Häh?

— Että hyvinkö se nukuttaa, uudisti kysyjä.

— Eihän tuo nyt oikein… On se välistä paremminkin…

— Onhan se ihanata, kun on ihmisellä hyvä unenlahja, virkkoi

Kantele.

Isäntä käsitti tämän pilan teoksi hänen makailuistaan ja korjaten

asentoaan sanoi:

— Omaa aikaanipa tässä makailen. Ei sen luulisi ketään

liikuttavan.

— Isäntä käsitti minut väärin, alkoi Kantele selitellä. Minä päin

vastoin tarkoitin, että unen lahja on hyvä olemassa ihmisellä.
 — Jaa, virkkoi isäntä vähän leppyneemmin ja aikoi kohta
kuorsahdella.

Kanteleen sisua alkoi kaivella moinen välinpitämättömyys isännän

puolelta, ja hän kysyi Villeltä otsa kurtussa.

— Enpä luule tässä talossa mitään tulevan. Eikähän tässä ole

ihmisiäkään muuta kuin tuo laiska isäntä.

— Kyllä tässä on väkirikas talo, mutta ovat ulkotöissä, selitti Ville.

Pitäisi tästä minun tietääkseni saada vakuutus. Leimauksen miehet
olivat käyneet joku päivä sitten ja olivat kehuneet tästä saavansa
ison vakuutuksen.

— Ahaa. Varsala se sitten on käynyt. Täällä onkin siis raudat

kuumana. Nyt taotaan pojat Mutta meidän pitäisi saada ruokaa.
Missä lienee emäntä?

— Näkyi se äsken liikkuvan tuolla ulkona…

Ville meni etsimään emäntää. Isäntä kuorsahteli jo niin että seinät

tärisivät, väliin muodostellen omituisia sihiseviä kurkkuääniä, jotka
kiusasivat ylen määrin Kanteletta.

— Perhanan ukko, kun ei välitä mitään vieraista. Se on sitä

talonpoikais-moukkamaisuutta, joka ei välitä, vaikka kuvernööri
taloon tulisi, vaan jatkaa inhoittavaa soitteluaan.

Emäntä löytyi sikolätistä ja tuli Villen kanssa tupaan, pyyhki

kätensä esiliinaan ja tervehti vierasta.

— Mitä sitä kuuluu? kysyi.

 — Eipähän tuota erikoisempaa, murahti Kantele, jota kyllästytti
semmoinen kyseleminen.

— En minä tunne tätä vierasta, jos lienee vaikka oman pitäjän

herroja, arveli emäntä.

— Ei tämä tästä pitäjästä… Tämä on Honkayhtiön tarkastaja,

selitti
Ville.

Emäntä vähän sävähti.

— Ja minä tässä vaan kaikenlaisia kyselen. Lähtekää vieraat

tuonne kamariin istumaan. Tuo ukkokin tuossa niin kuorsaa. Taitaisi
se kamari olla viileämpikin.

— Kyllä me täälläkin voidaan istua. Jos emäntä laittaisi sinne

kamariin meille ensin päivällistä.

— Eikö vieraat ole vielä sitäkään saaneet, hätääntyi emäntä ja

meni etsimään tyttäriään, jotka kohta ilmestyivätkin ruokapuuhiin,
kun ensin olivat saaneet pukeutuneiksi pyhävaatteisiinsa.

— Mitenkähän tuon isännän saisi ylös, virkkoi Kantele emännälle,

joka pistäytyi tupaan.

— Niinhän tuo raato äityi taas makaamaan.

Emäntä kävi tönimään aviotaan kylkeen.

— Nousetko siitä, retvana. No, kuuletko… kun on tullut vieraita ja

sinä syntinen vain makaat.

Ja vähän hiljemmin silmiään aukovan isännän korvaan:

 — Mikä lienee tarkastaja… jos lienee kuin iso herra.

— Häh?

— Että nouse nyt

— Nousenhan minä.

Ja isäntä vääntyi vaivaloisesti istumaan. Kopeloi sängyn laidan

alta tupakkavehkeet esiin ja alkoi latailla piippuaan. — Kovinpa se
nyt painostaakin. Eikö tulle ukonjyry, arveli isäntä haukotellen.

Kantele siirtyi lähemmä istumaan ja koetti rakentaa keskustelua.

— Tuleeko se nyt minkälainen heinävuosi? kysyi.

— Mistäpä tuon nyt vielä tietää, jurahti hetken perästä isäntä.

— Tässä näyttää olevan iso talo ja hyvämainen, aloitteli Kantele

— ja rakennuksetkin on erittäin hyvässä kunnossa. Tekee oikein
hyvää tulla tämmöiseen taloon, jossa on kaikki paikat reilassa.
Näkee heti, että talossa on järjestystä.

— Mistä päin sitä liikutaan? kysyi isäntä jo virkistyneenä,

— Tämä on Honkayhtiön tarkastaja, kiirehti Ville selittämään.

Tulemme sieltä meiltä viimeksi.

— Millä asioilla sitä liikutaan? kysyi isäntä, joka ei nähtävästi

muistanut sen nimistä yhtiötä.

— Olemme henkivakuutuksia päättämässä, ilmoitti Kantele. Tämä

naapurin mies tuli nyt tänne teidän kylälle asiamieheksi, ja minä
päätin lähteä häntä vähän opastamaan. Joko se on isännällä
henkivakuutus?

— Eipä sitä vielä…

— No nythän sopisi kuulla näitä meidän yhtiön ehtoja ja etuja,

jotka on jo nykyään tunnustettu parhaimmiksi, selitteli Kantele.
Kuulin, että täällä ovat käyneet Leimausyhtiön miehet tarjoamassa
vakuutusta; vaan eiköhän tuota nyt sentään näin ison talon isännälle
päätettäisi suuremmassa yhtiössä, ja kun ne ehdotkin ovat
paremmat.

— Kävihän niitä Leimauksen herroja täällä kiusaamassa.

— Minkälaista vakuutusta ne isännälle tarjosivat?

— En tuota tullut tarkemmin kuunnelleeksi, jurahti isäntä.

Kantele kaivoi taskustaan taulukon ja kysyi.

— Joko isäntä on vanha?

Isäntä vilkaisi syrjäkatsein vieraisiin.

— Joko ne peijoonit taas sitä ikää tutkii, mietti, mutta ääneen

virkkoi.

— Johan se viideskymmen lienee menossa.

— Jokohan tuo missä asti?

Emäntä joutui selittämään.

 — Se täyttää neljäkymmentäkolme vuotta siinä Marjanpäivän
aikana.

Sitäkö henkivakuutusta se tämäkin herra on tarjoomassa?

Saisihan se ukko siihen vakuuteen ruveta, vaan sepä onkin
semmoinen pässinpää, ettei uskalla. Pelkää rahojensa menevän sitä
tietään.

— Älähän tuossa taas, kivahti isäntä. — Se on tuo meidän akka

aina suuna ja päänä.

— Emäntä ymmärtää vakuutuksen merkityksen, ja ymmärtää sen

isäntäkin, kun vaan tekee kerran ratkaisevan päätöksen. Meidän
yhtiö antaa suurimman varmuuden vakuutetuille.

— Ottaisihan sen vakuutuksen, mutta sattuisi vielä tässä

kuolemaan, ja leski pääsisi reuhaamaan niillä rahoilla. Menisi vielä
uusiin naimisiin ja naureksien nauttisivat mitä toinen on säästänyt,
arveli isäntä.

— Tuosta sen kuulette vieraat, minkälainen vätys se on. Pelkää

perillistensä rikastuvan. Ottaisit vain sen vakuutuksen.

— Menehän tuosta, murahti isäntä ja kellahti uudelleen sänkyyn.

— Täällä olisi ruokaa vieraille, ilmoitti emäntä.

Kantele kumminkaan ei joutanut nyt ajattelemankaan ruokaa.

Hänestä näytti hyvin mahdolliselta vakuutuksen päättäminen, ja
selitettyhän yhtiönsä etuja kävi hän sängyn viereen, jossa isäntä
silmät puoliummessa veteli haikuja piipustaan, ja virkkoi:

— Mutta ajatelkaapas, jos kuolema tulee kuinka pian.

 — Kun tulee, niin tulee. Mitäpä hänestä…

— Ette siis ajattele lastenne tulevaisuutta.

— Tottapahan eteensä katsovat. Ja jäähän tästä vähän tätäkin

taloriekaletta.

— Kyllä jää. Komia talo jääkin, vaan olisihan se kymmenen tuhatta

lisänä.

— Mitäpä ne niin suurilla rahoilla.

Isäntä alkoi kuorsahteiemaan.

— Miettikäähän asiata tarkemmin, kehoitteli Kantele.

— Joo… puh… huu…

Isäntä kuorsaili jo harvassa tempossa, ja Kantele meni Villen

kanssa kamariin.

— Tässä ei nyt ehditty laittamaan oikein herroiksi, selitteli emäntä.

Olisihan sitä osattu laittaa, kun meidän Hanna on käynyt
keittokoulunkin, Joko sitä herroilla on rouvaa?

— Ei ole.

— Vai ei. Ovatpahan malttaneet olla, ihmetteli emäntä ovessa

mennessään.

— Koko pöljä sentään tuo isäntä, murisi Kantele. Luulin jo

vakuutuksen päättyvän, vaan ukolla ei ole nähtävästi mitään muuta
harrastusta, kuin että saa vahvasti kuorsata.
 Ville oli käynyt talon palvelusväelle puhumassa Honkayhtiön
suurista eduista ja kuullut, että Pörjönen, Djefvulsund yhtiön
tarkastaja oli käynyt asiamiehineen naapuritalossa ja vakuuttanut
siellä palvelijat ja myöskin Kuivalan rengin. Kuulemansa uutisen
kertoi hän Kanteleelle, joka tästä vimmastui.

— Mikä pahuuksen Pörjönen se on, joka käy noin vain nenän

edestä nappaamassa vakuutukset, ja me saamme nuolla
näppiämme. Nähtävästi siitä yhtiöstä koituu meille vaarallinen
kilpailija.

— Heillä on syöttinä se työkyvyttömyysvakuutus, tiesi Ville. Tässä

Honkayhtiössä ei taida sitä ollakkaan?

— Se on tarpeeton. Sen nojalla saavat vain kantaa korkeammat

vuosimaksut. Voihan se vastaisuudessa tulla meidänkin yhtiöömme,
vaan siihen mennessä täytyy nyt Villenkin muistaa leimata se
kokonaan keinotteluksi. Ymmärtääkö Ville?

— Kyllä.

— Ja aina pitää muistuttaa vakuutettaville, että Honkayhtiö on

vakavaraisin, neuvoi Kantele.

— Samaa ne kuuluu väittävän Riento yhtiönkin miehet, kertoi Ville.

— Jaa, mutta se pitää leimata heti valheeksi.

Henkivakuutusasiamiehen täytyy osata tarpeen tullen valehdella ja
osata näyttää, että kilpailevan yhtiön mies on pannut omiansa.
Hyvän asian edesauttamiseksi on lupa käyttää valhettakin. Ja kansa
ei usko vakuutuksen hyviä etuja, jollei niitä tuntuvassa määrässä
lisäillä.