Medical Image Reconstruction

Term II – 2012

Topic 3: Image Reconstruction

in the Fourier Domain

Professor Yasser Mostafa Kadah

Topics Today
 Reconstruction from Uniformly Sampled k-Space
 Projection-slice theorem
 Interlaced Fourier transform
 Partial Fourier Methods
 Reconstruction from Nonuniformly Sampled k-Space
Reconstruction from k-Space Samples
 Several imaging modalities produce data in k-space
 Computed Tomography (CT): radial sampling
 Magnetic Resonance Imaging (MRI): several sampling strategies
are used such as radial, spiral, and random sampling
 Main Reconstruction Method: Compute Inverse 2D DFT
to compute the image!
Image Reconstruction Problem

Sampled k-Space Image Spatial Image

Reconstruction
System

Sampling Pattern
Projection-Slice Theorem
 Also known as Central-Slice Theorem
 A property of the Fourier transform
 Relates the projection data in the spatial domain to the
frequency domain
 States that the 1D Fourier transform of the projection of
an image at an angle  is equal to the slice of the 2D
Fourier transform at the same angle
Projection-Slice Theorem
y ky

2D
f(x,y) F(kx,ky)
Fourier
Transform
x 

 kx


1D
Fourier
p () Transform
P (k)
p ()

k
Projection-Slice Theorem
 2D Fourier transformation:
 j 2 ( k x  x  k y  y )
F (k x , k y )   f ( x, y)  e dxdy
 The slice of the 2D Fourier transform at kx=0 is given by:

F (0, k y )    f ( x, y)dxe  j 2k y  y

dy
and at ky=0 is given by

F (k x ,0)    f ( x, y)dye  j 2k x  x

dx
Projection-Slice Theorem
 For a general angle, the rotation property of the Fourier
transformation can be used to generalize the
mathematical result for a vertical projection to any angle
0 45 90
Projection-Slice Theorem:
Application to CT
 The projection data can be shown to correspond to
radial sampling of the frequency domain
 It is not straightforward to numerically compute the
image from this frequency domain representation
 Limitation of the DFT to uniform sampled data
 Interpolation can be used in the frequency domain to re-
grid the radial sampling to uniform sampling
Projection-Slice Theorem:
Application to MRI
 Navigator echo motion estimation
 Acquire a single k-space line in the middle to estimation linear
translation in this direction
 Early MRI reconstruction based on backprojection
algorithms
Interlaced Fourier Transform
 A special case of nonuniform Fourier transform
Interlaced Fourier Transform
 Mathematical formulation
Interlaced Fourier Transform
Shepp-Logan Phantom
 Numerical phantom used to simulate the human head to
evaluate reconstruction algorithms in computed
tomography
Shepp-Logan Phantom
Shepp-Logan Phantom: 3D
Shepp-Logan Phantom: k-Space
 Using the known Fourier transformation of the Shepp-
Logan phantom components (circles and ellipses), one can
generate the analytical form of its Fourier transformation
 Can be sampled arbitrarily to generate uniformly or
nonuniformly sampled data for close to real data generation
 Applications include radial sampling (e.g., CT and MRI), spiral
and random sampling (MRI).
 This will be the standard for all evaluation procedures of
image reconstruction methods.
Partial Fourier Reconstruction
 PF reconstruction is based on the fact that if the object is
real in image space, its Fourier transform is Hermitian.
 One-half of the k-space is needed to reconstruct a real image
 In reality, however, the reconstructed images are complex.
Partial Fourier
Reconsruction
Conjugate Synthesis
 Assume image is purely real
Conjugate
Synthesis
Margosian Method
Cuppen/POCS Method
Phase Correction Effects
FIR and MoFIR Methods
Comparison of Distortions in
Margosian and MoFIR Methods
Experimental Verification:
Simulated Box and Real Data
Results
Results
Results
Results
 256x256 image, 16 Lines

 8 Lines
Reconstruction from Nonuniformly
Sampled k-space: Conventional Gridding
 Conventional gridding through convolution with
interpolation kernel (O’Sullivan, 1985; Jackson et al., 1991;
Meyer et al., 1992)
Selection of
Gridding
Kernel
Sampling
Trajectory
2D Sampling Inverse 2D
Density
Convolution Compensation FFT
Acquired
k-space
Data Apodization
Removal

Image
Meyer Gridding Algorithm
Meyer Gridding Algorithm

Nonuniform Samples

1x Grid

2x Grid
Jackson Gridding Window Selection
Jackson’s Gridding Window Selection
Disadvantages of Gridding Methods
 Reconstructed images do not represent optimality in any
sense
 Variation of performance with form of k-space trajectory
 Lack of explicit methodology to trade-off accuracy and
speed of reconstruction
 Not possible to progressively improve the accuracy of
reconstruction
Kadah’s Method

 Algebraic Solution
 Iterative reconstruction method that provides an optimal
solution in the least-squares sense
 Based on a practical imaging model
 Progressive reconstruction capability
 Simple mechanism to control trade-off between accuracy
and speed
Theory
 Assume a piecewise constant spatial domain representing
display using pixels
 Image composed of pixel each of uniform intensity
 Image can be represented by a sum of 2D RECT functions
 Assume spatial domain to be compact
 Field of view is always finite in length
 The image can be expressed in terms of gate functions as,
N 1 M 1
f ( x, y )   n,m   ( x  xn , y  y m )
n 0 m 0
 Theory
 Applying continuous Fourier transform,
  N 1 M 1

   
 j 2 ( k x x  k y y )
F (k x , k y )  n,m   ( x  xn , y  ym )  e dxdy ,
   n  0 m  0

 Hence,
N 1 M 1
F (k x , k y )  Sinc ( wx k x )  Sinc ( w y k y ) n,m  e
 j 2 ( k x xn  k y ym )
.
n 0 m 0

 This can be expressed in the form of a linear system as

b  Av
 A matrix is ~N2N2 and complex-valued
Theory
 Observation: A matrix is ~N2xN2 and complex-valued
 Solve a16384x16384 linear system to get a 128x128 image
 Very difficult to solve in practice because of size
Idea
 Problem: A matrix is dense and computational
complexity of solution is prohibitive
 Solution Strategy: Try to make the A matrix sparse by
seeking a compact representation of rows in terms of
suitable basis functions
 Observation: applying a 1-D Fourier transformation to
the rows of A matrix results in energy concentration in
only a few elements
Methods
 Multiply the rows of the system matrix by the NxM-point
discrete Fourier transform matrix H in the following
form:
 Methods
 How to multiply H without changing the linear system?
 Row energy compacting transformation converts the system into a
sparse linear system as follows:

b  Av  A  H H  H  v  (H  A H ) H  V  M  V ,
 To convert to sparse form, only a percentage  of kernel
energy in each row is retained
 The only parameter in the new method
 Correlates directly to both image quality and computational
complexity
 Sparse matrix techniques are used to store and manipulate the
new linear system
 Since the linear system is sparse, iterative methods such as conjugate
gradient can be used to solve the system with very low complexity
Methods
Results
 256x256
Analytical
Shepp-Logan
Phantom
(Radial sampling)
Results
 256x256 Real data from a resolution phantom at 3T from
a Siemens Magnetom Trio system using a spiral trajectory
Results
Discussion
 Full control over the accuracy versus complexity trade-off
through  selection
 Computational complexity is comparable to conventional
gridding with small kernel
 O(g()L) per CGM step, where g() is the average # of
elements/row, L=# of acquired k-space samples
 Average 4.9 elements/row to retain 92% of energy
 Progressive reconstruction is possible
 Add more iterations to process
 Use a different reconstruction table with higher 
Exercise
1. Use the MRI data set on the web site and write a
program that reconstructs the image using a 2D inverse
Fourier transform.
2. Write a program to verify the projection-slice theorem
using a simple 2D phantom (e.g., a basic shape like a
square).
3. Perform interlaced sampling on a function of your
choice with known analytical Fourier transform and
verify the interlaced Fourier transform theorem.
4. Write a Matlab program to implement the analytical
Shepp-Logan phantom and test it using sampling on a
uniform grid.
Exercise
5. Write a short paragraph (less than 500 words) on
which partial Fourier reconstruction method you prefer
and why.
6. Use the data set on the class web site to implement
one of the methods of partial Fourier reconstruction.
The data set provided is for full k-space for you to have
a gold standard to your reconstruction.You should use
only part of it as an input to your reconstruction (say
half + 16 lines).
7. Do a literature search on the topic of partial Fourier
reconstruction and come up with a list of all references
related to the subject.
Exercise
8. Do a literature search on the problem of nonuniform
sampling in 2D and summarize your findings about the
sampling criteria to avoid aliasing in less than 500 words
(in addition to a list of references).
9. Write a program to perform gridding on generated
radial k-space sampling of the k-space of the numerical
Shepp-Logan phantom to compute the image.
10. In less than 500 words, describe how one can compare
the quality of different reconstruction methods and/or
parameters based on measurements from the generated
images.
Exercise
11. Verify the energy compactness transformation and
generate in Kadah’s method for any trajectory you
prefer.
12. Assuming that we have a rectilinear sampling instead of
the nonuniform sampling in this paper, how do you
expect the linear system to look like?
13. Assume that we are constructing an NxN image,
compute the exact number of computation (not an
order or computation) detailing the list of computations
in each step in the implementation.