Download as pdf or txt
Download as pdf or txt
You are on page 1of 30

COMPRESSED SENSING BASED REAL-TIME DYNAMIC MRI

RECONSTRUCTION
Angshul Majumdar, Rabab K. Ward and Tyseer Aboulnasr
Department of Electrical and Computer Engineering, University of British Columbia
angshulm@ece.ubc.ca, rababw@ece.ubc.ca and taboulnasr@apsc.ubc.ca
Abstract –
This work addresses the problem of real-time online reconstruction of dynamic MRI sequences. The
proposed method reconstructs the difference between the previous and the current image frames. This
difference image is sparse. We recover the sparse difference image from its partial k-space scans by using
a non-convex Compressed Sensing algorithm. As there was no previous fast enough algorithm for real-
time reconstruction, we derive a novel algorithm for this purpose. Our proposed method has been
compared against state-of-the-art offline and online reconstruction methods. The accuracy of the proposed
method is less than offline methods but noticeably higher than the online techniques. For real-time
reconstruction we are also concerned about the reconstruction speed. Our method is capable of
reconstructing 128 x 128 images at the rate of 6 frames per second, 180 x 180 images at the rate of 5
frames per second and 256 x 256 images at the rate of 2.5 frames per second.
1. INTRODUCTION
This paper aims at real-time online reconstruction of dynamic Magnetic Resonance Image sequences.
There are broadly two classes of dynamic MRI reconstruction methods – offline and online. Offline
methods reconstruct the images after all the data (pertaining to all time frames) have been acquired.
Online methods reconstruct the images for each time frame individually. Online methods are not
necessarily real-time. Real-time reconstruction is a special case of online reconstruction, where the
images (corresponding to each time frame) are reconstructed as fast as they are acquired. Real-time MRI
will benefit several clinical applications including cardiac cine MRI, MR image guided therapy and
catheter tracking.
For dynamic MRI, the k-space samples of each time frame are collected; the raw data are therefore in the
k-t space. The problem is to reconstruct each time frame, so that the final outcome is an MRI video (x-t
space). The acquisition speed of the MRI scanner is limited by its hardware physics; in general there is a
trade-off in spatial and temporal resolution, i.e. if the number of k-space samples collected for each time
frame are large then the number of time frames that can be acquired in a unit of time is small and vice
versa.
Previously, the k-space had to be densely sampled on a uniform Cartesian grid and the images for each
time frame were reconstructed by applying the inverse Fourier Transform. Thus, in order to achieve high
spatial resolution, temporal resolution had to be sacrificed. There are also techniques to sample the k-
space efficiently on a non-Cartesian grid. Reconstruction from non-Cartesian samples is not straight-
forward; the image needs to be obtained iteratively (typically via least squares minimization). Such non-
Cartesian sampling required over-sampling of the k-space in order to yield images of good quality. In
recent times, Compressed Sensing based techniques [1-4] have enabled the reconstruction of MR images
from partially sampled k-space scans. Therefore it is now possible to acquire less number of k-space
samples for each time frame and reconstruct them without sacrificing spatial resolution. The number of
samples to be collected for each frame has thus been reduced thereby improving temporal resolution
without sacrificing spatial resolution.
Most previous studies in Compressed Sensing based dynamic MRI reconstruction employed offline
methods [3-6]. Thus, the reconstruction speed was not of concern. They applied sophisticated
Compressed Sensing based reconstruction methods that exploit both spatial and temporal redundancies of
the whole dynamic MRI dataset.
The focus of this work is on real-time reconstruction. Ideally one would expect to obtain dynamic MRI
sequences with high spatial and temporal resolution. But, owing to the physical limitations of the MRI
scanner, high temporal resolution can only be achieved, by under-sampling the k-space for each time
frame. This poses a challenging task for real-time reconstruction. On one end, high temporal resolution
requires that the images for each time frame has to be reconstructed very fast (as fast as they are
acquired). This precludes the use of sophisticated reconstruction methods which are generally time
consuming. On the other end, reconstructing individual frames from highly under-sampled k-space is a
challenge even for Compressed Sensing based methods.
In this work, we propose a very fast scheme for dynamic MRI reconstruction that can achieve near real-
time performance. Instead of reconstructing the full image at each time frame, we reconstruct the
difference image between the previous time frame and the current one. Such temporal differencing
schemes have been used previously for offline reconstruction [7-11]. The aforementioned studies assume
that the motion of major parts of the anatomy under the scanner is slow, and only a few portions have
rapid motion. Thus, under temporal differencing of two consecutive time frames, areas with slow motion
are cancelled (zeroed) out (approximately) and only areas with high motion contribute to the difference
image. This leads to a sparse difference image. Using this assumption, the aforesaid studies were able to
produce very good offline reconstruction results. In this work, we address the problem of online
reconstruction by using the same assumption used earlier [7-11], i.e. the difference image between two
consecutive time frames is sparse.
This assumption makes the problem of reconstructing the difference image an ideal candidate for
Compressed Sensing (CS) recovery. The signal (difference image) to be reconstructed is sparse (i.e. it is
sparse in Identity/Dirac basis); the k-space data is acquired in the Fourier domain. Thus the measurement
basis (Fourier) and the sparsifying basis (Idenitity/Dirac) are maximally incoherent among all possible
basis pairs. This incoherence property of the problem makes it suitable for CS recovery.
Instead of using the standard convex l1-norm minimization which is profusely employed in CS recovery,
we employ a non-convex lp-norm minimization. Previous studies in MRI reconstruction showed that non-
convex optimization yield better reconstruction results than their convex counterparts [2, 12-16].
However, all lp-norm minimization algorithms are unfortunately slow. They are based on 1) the iterative
reweighted least squares method (also called the FOCally Under-determined System Solver - FOCUSS)
[13] or 2) the iterative reweighting of l1-norm minimization [15, 16] or 3) the homotopy approaches
which solves the non-convex problem through a continuation scheme [2, 12]. The speed of the
reconstruction algorithm has not been an issue for previous studies as they were reconstructing MR
images offline. In this work, we want to reconstruct the dynamic MR images in real-time, therefore our
reconstruction algorithm has to be very fast. Since no off-the-shelf non-convex reconstruction algorithm
meets the demands of real-time reconstruction, we derive a fast shrinkage algorithm to directly solve the
lp-norm minimization problem.
We have carried out experimental validation on several real dynamic MRI datasets of varying temporal
and spatial resolution. We have compared our method with state-of-the-art offline [5, 6] and online [18,
30] reconstruction methods. The reconstruction accuracy from our proposed method is slightly worse than
the offline methods, but better than the online methods we compared against. Our proposed method can
reconstruct small images at the rate of 5-6 frames per second and large images at the rate of 2.5 frames
per second. This is near real-time speed for dynamic MRI.
The rest of the paper is organized into several sections. The following section briefly discusses the basics
to Compressed Sensing. Related previous studies in offline and online dynamic MRI reconstruction are
reviewed in Section 3. Section 4 details our proposed approach and derives the reconstruction algorithm.
The experimental results are given in Section 5. The conclusions of this work are discussed in Section 6.
2. AN OVERVIEW OF COMPRESSED SENSING
Compressed Sensing (CS) is concerned with the solution of under-determined inverse problems when the
solution is known to be sparse.
ym1  Amn xn1  m1 , m  n (1)

In signal processing terminology, y is the vector comprised of the collected samples and A is the
measurement operator/matrix (for MRI it is the restricted/masked Fourier Transform). The signal to be
recovered is x and is assumed to be s-sparse, i.e. it has s non-zero coefficients while the rest n-s are
zeroes. The problem is supposed to be corrupted by white Gaussian noise  N (0, 2 ) .

This is an under-determined inverse problem having infinite number of solutions. CS studies the
conditions under which the inverse problem can be uniquely solved by practical recovery algorithms.
The signal to be recovered x is assumed to be s-sparse. An s-sparse vector has 2s unknowns – its s
positions and the corresponding values at each position. Therefore, if the number of samples (m) is larger
than 2s, it will be possible to recover x by brute force search. This is an NP hard combinatorial
optimization problem that can be represented as,

subject to y  Ax 2  
2
min x 0
(2)
x

where the x 0 is the cardinality of the vector and   n   2 .

There is no feasible algorithm to solve (2) in polynomial time. In order to alleviate this problem, seminal
papers in Compressed Sensing [22, 23] proposed substituting the NP hard l0-norm by its closest convex
surrogate the l1-norm. Thus they proposed a solution to (1) via the following optimization,

min x 1 subject to y  Ax 2  
2
(3)
x

This is a convex optimization problem, and there are efficient algorithms to solve it. However, the
number of samples m required to guarantee signal recovery for (3) is greater than that for (2); the number
of samples required are,
m  Cs log n, C is a constant (4)
In between the two extremes of NP hard optimization (2) and convex optimization (3) there lies non-
convex optimization problems of the following form,

subject to y  Ax 2   , 0  p  1
p 2
min x p
(5)
x

The number of samples required for solving (5) to succeed is,


m  C1s  pC2 s log n, C1 and C2 are constants (6)

For high value of p, the problem (5) is close to the convex problem (3) and the sampling requirement in
(6) is dominated by the term pC2 s log n , which is similar to the requirements of the l1-norm minimization.
But when the value of p is small, (5) is close to the NP hard problem (2); also the term dominating the
sampling requirements in (6) increases linearly with the sparsity ( C1s ) which is similar to the requirement
of the l0-norm minimization. In summary, as we decrease the value of p, the number of samples required
will also decrease for a fixed value of s. In other words, if the number of samples (m) is fixed then non-
convex optimization can recover an s-sparse vector that have higher value of s than that possible by
convex optimization. But the disadvantage of non-convex minimization is that the optimization algorithm
may be terminated in a local minima and may fail to reach the desired solution. However, in practical
MRI reconstruction problems [2, 12-16] this had never been the case.
Practical signals are not exactly sparse. They are only approximately sparse (have a power law decay of
index q); but we model them to be exactly sparse in order to reconstruct them by the optimization
algorithms mentioned above. Let us assume that x0 is the (approximately) sparse original signal that is to
be recovered; and let xˆs be the s-sparse signal recovered by CS recovery algorithms; then x0  xˆs is the
modeling error. The upper bound on the modeling error is inversely proportional to s [23],

x0  xˆs  Cs s ( q 1/2) (7)

In real-life problems, it is not possible to 1) know the index of the power law decay ‘q’, 2) estimate the
allowable modeling error, 3) estimate the number of sparse coefficients to be recovered and 4) decide the
number of samples (m) to be collected. In practice the solution proceeds exactly in the opposite fashion.
The number of collected samples is fixed (in our case by the hardware of the MRI scanner and spatio-
temporal resolution of the dynamic MRI sequence), and one needs to reduce the modeling/reconstruction
error to the maximum possible extent. Experimental studies in MRI reconstruction [2, 12-16] show that
non-convex lp-norm minimization yields better reconstruction results than its convex l1-norm counterpart.
Most practical signals (MR images, seismic images and natural images) are not sparse themselves. They
have a sparse representation in a transform domain such as wavelets, curvelets or Discrete Cosine
Transform (DCT). These transforms are either orthogonal or tight-frames and hence can be represented
by the analysis-synthesis pair of equations,
Analysis:   x (8a)

Synthesis: x  T  (8b)

where Ψ is the sparsifying transform (wavelets, curvelets, DCT etc.) and α is the approximately sparse
transform coefficient vector.
When the signal to be recovered is sparse in a transform domain, the inverse problem (1) is modified as,

y  AT    (9)

Thus, the following optimization problem needs to be solved in order to recover α,


2
min  subject to y  AT  
p
p
(10)
x 2

when p=1, (10) represents the convex l1-norm minimization and when 0<p<1 it represents the non-
convex lp-norm minimization.
Once the sparse coefficients α are recovered, the signal is reconstructed by applying the synthesis
equation (8b). This standard CS based technique for MRI reconstruction (equations 7-10) has been used
before in [1, 2, 12-14].
The number of samples required to reconstruct the signal by l1-norm minimization is [23],
m  C  s log n (11)

where µ is the mutual coherence, defined as the maximum of the off-diagonal elements of the Gram
matrix AT AT .
Among all possible basis pairs, the mutual coherence is the least between the Fourier and Dirac basis, i.e.
when A is a Fourier basis and Ψ is a Dirac basis [23]. For any other pair of basis (such as Fourier and
Wavelet – used in most CS based static MRI reconstruction), the value of mutual coherence is higher.
This means (loosely speaking) that for a given number of samples, the number of sparse coefficients that
will be recovered by CS based techniques will be the highest for Fourier-Dirac basis pair. In other words,
given a fixed number of samples, the Fourier-Dirac basis pair will result in the least
modeling/reconstruction error (7) for any CS recovery method.
As will be shown later, for this particular work, the signal to be reconstructed (the difference frame) is
sparse by itself; therefore our measurement basis is Fourier (k-space) and the signal is sparse in Dirac
basis. Hence, this is the best possible situation in terms of CS based recovery. However, for most practical
problems in MRI, the signal (MR image) is not sparse in itself but have a sparse representation in a
transform domain (such as wavelets). In that case, the number of samples required for recovering the
image increases owing to a larger μ.
3. LITERATURE REVIEW
Dynamic MRI reconstruction is an active area of research and there are a lot of papers on this subject. We
will only cover some papers that are relevant to our work.
3.1 Offline Reconstruction
The dynamic MRI data acquisition model can be expressed succinctly. Let Xt, denote the MR image
frame at the tth instant. We assume that the images are of size N x N and T is the total number of frames
collected. Let yt be the k-space data for the tth frame. The problem is to recover all Xt’s (t=1…T) from the
collected k-space data yt’s. The MR imaging equation for each frame is as follows,
yt  Rt F2 D xt   (12)
where F2D is the 2D Fourier operator which maps the image space to the k-space, Rt is the under-sampling
mask on the k-space, xt is the vectorized MR image formed by row/column concatenation of the image
marix Xt and η is white Gaussian noise.
The CS approach [3, 25] employs the wavelet transform to sparsifying the spatial redundancies in the
image and the Fourier transform for sparsifying along the temporal direction. This is a realistic approach
because both of them work with dynamic imaging of the heart where the change over time is quasi-
periodic, thereby leading to a compact (sparse) support in the frequency space.
The model from (12) can be organized in the following manner,
y  RF2D x   (13)

 y1   R1 ... 0   x1 
where y   ...  , R   ... ... ...  and x   ... 
   
 yT   0 ... RT   xT 

In both [3, 25] the standard CS optimization problem was proposed to reconstruct the MR image data x
from the k-space samples y,

min W2 D  F1D ( x) 1 subject to y  RF2 D x 2  


2
(14)
x

where W2 D  F1D is the Kronecker product1 of the 2D wavelet transform (for sparsifying in space) and F1D
the 1D Fourier transform (for sparsifying along the temporal direction). In the original studies [3, 25] the
Kronecker product is not used; we introduce it here to make the notation more compact.
The 1D Fourier transform may not always be the ideal choice for sparsifying along the temporal direction
(if the signal is not periodic); in such a case the 1D wavelet transform can be used instead. Some other
studies [7-9] assumed the image to be sparse in the x-f space, i.e. the signal was assumed to be sparse in
I 2 D  F1D basis (where I2D is the Dirac basis). These studies emphasized on using the FOCUSS [26]
method to solve (14). However, in practice any other state-of-the-art l1-norm minimization solver (e.g.
Spectral Projected Gradient L1 or Nesterov’s Algorithm) will work.
Some other studies [4, 5] either assumed the MR image frames to be spatially sparse or were only
interested in the ‘change’ between the successive frames. These studies did not explicitly exploit the
redundancy of the MRI frames in the spatial domain, rather they only applied a Total Variation (TV)
regularization in the temporal domain. The optimization problem they addressed is the following,

min TVt ( x) subject to y  RF2 D x 2  


2
(15)
x

N2
where TVt   || t xi || and  t denotes the temporal differentiation for the ith pixel.
i 1

 a11 ... a1n   a11 B ... a1n B 


1   
Let Amn and B pq be two matrices and A  ... ... ... , then A  B  ... ... ... 
  
 am1 ... amn   am1 B ... amn B 
In essence, the approach in [4, 5] differs only slightly from [7-9]. In [7-9] it is assumed that the signal
varies smoothly over time, so that it is sparse in the temporal Fourier transform. In [4, 5], temporal
Fourier transform is not directly employed, but the signal is assumed to be smoothly varying with time
with only a finite number of discontinuities, so that it will be sparse in the temporal differencing domain.
There is yet another class of methods that reconstruct the dynamic MRI sequence as a rank deficient
matrix. The k-space acquisition model is the same as (12), but instead of concatenating the data as
columns in (13), the data is stacked as columns in the following manner,
Y  RF2D X   (16)

where Y   y1 | ...| yT  , X   x1 | ...| xT  and   1 | ...| T  .

In [6, 27] it is argued that the matrix X is rank deficient, since it can be modeled as a linear combination
of very few temporal basis functions. Based on this assumption, [26] proposed solving the inverse
problem (16) as follows,

min X * subject to Y  RF2 D X 


2
F
(17)
X

where . * denotes the nuclear norm of the matrix and . F denotes the Frobenius norm of the matrix.

For recovering a rank deficient solution, one ideally has to minimize the rank of the matrix subject to data
constraints. However, minimizing the rank is an NP hard problem similar to l0-norm minimization. Thus,
the research in this area have proposed substituting the NP hard rank by its tightest convex surrogate the
nuclear norm [28, 29]. Following these studies, the inverse problem (16) is solved via a nuclear norm
minimization (17).
A related work in this area [6] proposed combining the rank deficiency of the signal in the x-t space
(similar to [26]) along with its sparsity in the spatio-temporal differencing domain (similar to [6]). The
following optimization problem was proposed to solve the dynamic sequence

min X *   S  t ( x) 1 subject to Y  RF2 D X 


2
F
(18)
X

where  S is the spatial differencing operator and  t is the temporal differencing operator.

Here X denotes the signal in the x-t space as arranged in [16] whereas x denotes the signal in the x-t space
as arranged in [13]. The term λ balances the relative importance of rank deficiency and sparsity.
3.2 Online Reconstruction
All the studies discussed above are off-line reconstruction techniques. Online reconstruction techniques
reconstruct the image corresponding to the current frame, given the reconstructed images till the previous
time frame. In applications of dynamic MRI such as MRI guided surgery the reconstruction has to be
performed in real-time. Online reconstruction does not guarantee a real-time solution. To be real-time, a
reconstruction technique should be able to recover the images as fast as the frames are acquired.
Real-time dynamic MRI reconstruction is a new area of research and there is only a handful of papers on
this subject. In recent studies [18-20] real-time dynamic MR reconstruction algorithms that are based on
Kalman filtering are proposed.
In [18, 19], MRI reconstruction is formulated as a dynamical system; the following the Kalman filter
model is proposed,
xt  xt 1  ut (19a)

yt  Rt F2 D xt  t (19b)

where the pixel values xt form the state variable, the k-space sample yt represents the observation, ηt is the
observation noise and ut is the innovation in state variables.
In general, the Kalman filter is computationally intensive since it requires explicit matrix inversion for
computing the covariance matrix. However, in [18, 19], it has been shown that this problem can be
alleviated by diagonalizing the covariance matrix. To be able to do so, the pixel values of the difference
image must be uncorrelated. This is not a realistic assumption for MRI (or any natural motion in general).
In [20, 21], Kalman filtering of the wavelet coefficients is proposed. The filter model is the following,
t  t 1  ut (20a)

yt  Rt F2 DW2TDt  t (20b)

where  t is the wavelet transform coefficient vector for the tth frame.

This work [20] was motivated by the findings of compressed sensing. It proposed a further check over the
Kalman update, i.e. when the error between the prediction and the actual data becomes large, instead of a
filter update, full reconstruction using CS technique is done.
In recent work called LS-CS [17], it was proposed that instead of solving the image at the current time
frame, the difference between the previous frame and the current one can be reconstructed instead. The
following optimization problem is proposed,
2
min t subject to yt  Rt F2 D xt 1  Rt F2 DW2TD t  (21)
t 1 2

where xt 1 is the reconstructed image for the previous frame and  t is the wavelet transform of the
difference between the previous and the current frame.
The main assumption of [17] is that the sparsity pattern varies slowly with time. For each time frame, the
difference of the wavelet transform coefficients  t is computed, from which an intermediate estimate of
the wavelet coefficients of current frame is obtained as,
t  t 1  t (22)

The intermediate estimate is thresholded (via hard thresholding) and all the values below a certain
threshold τ are discarded. The value of the thresholded wavelet coefficients are not of importance, but
their indices are. Assuming that the set of indices having non-zero values after thresholding is Ω, the final
value of the wavelet transform coefficient is estimated as,

t  ( Rt F2 DW2 D )† yt (23)


where ( )† denotes the Moore-Penrose pseudo-inverse and ( ) indicates that only those columns of the
matrix has been chosen that are indexed in Ω.
The image is reconstructed from the wavelet coefficients via the synthesis equation.
Another recent work on dynamic MRI reconstruction proposed a method called Modified CS [30]. This
study assumes that in dynamic MRI, the support of the MR images evolve slowly with time; therefore it is
possible to estimate a portion of the support for the current frame given information of the support from
the previous frames. Modified CS estimates the known part of the support (Γ) of the sparse transform
coefficients for the current frame, given the previous reconstructed frame. Mathematically, the
optimization problem to be solved is,
2
min t(  ) subject to yt  Rt F2 DW2TD t 
t 1 2

where  t(  ) is the sparsest solution containing the support (Γ).

4. PROPOSED METHOD
In this paper we follow a reconstruction method similar to [17], i.e. instead of reconstructing the full
image at the current time frame, we reconstruct the difference between the current one and the previous
one. The formulation and the solution to our problem are however different from the aforementioned
work. In [17] it is assumed that the difference image (between current and previous frame) is not sparse
by itself but is sparse in the wavelet transform domain. This follows from studies in video coding and
compression, where the residual between the frames are encoded using some transform like DCT or
wavelets.
In dynamic MRI, the difference image between two consecutive time frames is itself sparse. This
assumption is implicitly used in [7-9] for formulating the reconstruction algorithm. In [4, 5] the
assumption is used explicitly for sparsifying the dynamic MRI sequence along the temporal direction. Our
work is based on the same assumption. We express the image at the current time frame as the sum of the
previous image and a difference.
xt  xt 1  xt (24)

whereas [17] assumed that the difference xt is sparse in a transform domain. In this work, we assume that
xt is sparse in itself. This is true for dynamic MRI sequences because the major portion of the anatomy
under study is static or is very slowly moving, while only small areas have significant movement. Thus
when the difference between consecutive frames is taken, the static and slowly moving areas cancel out
making the corresponding areas of the difference image filled with zeros. Only the fast moving portions
contribute to the non-zero values in the difference image. Thus the difference image xt is itself sparse in
the pixel domain. Notice that contrary to Kalman Filter based techniques [18-21], we do not make any
assumption about the model of xt .

The empirical verification on three consecutive frames from a larynx sequence is shown in Fig. 1. It
shows the difference images of two consecutive frames. The decay of the pixel values and the wavelet
coefficients (Daubechies wavelets with 4 vanishing moments, at 3 levels of decomposition) for the
difference images are shown in Fig. 2. The decay curve is the plot of the sorted pixel values or wavelet
coefficients. The decay curves give an estimate of the approximate sparsity of the signal. If the decay is
fast, most of the signal energy would be concentrated in few coefficients.

Fig. 1. Top: Three consecutive images from a larynx sequence; Bottom: Difference images between the three consecutive frames
Fig. 2. Top: Sorted pixel values of the difference images; Bottom: Sorted wavelet coefficients of the difference image.

It can be seen from Fig. 2 that the both the pixel values of the difference image and the wavelet transform
coefficients of the difference image has fast decays, i.e. most of the signal energy is confined in a
relatively small number of high valued coefficients. The energy confined in the top 10% of the pixels in
the difference image is 93% of the total energy and in top 10% of the wavelet coefficients it is 98%. Thus
the difference image has a slightly sparser representation in the wavelet domain than in pixel domain.
However, the slim relative advantage of wavelets in terms of energy compaction in fewer coefficients will
be offset by other factors as will be discussed now.
The fact that mutual coherence between the random Fourier ensembles and Dirac basis is the minimum is
well known in CS based MR image reconstruction literature [31, 32]. Independent of the k-space (Fourier
domain) under-sampling ratio, the mutual coherence is always ‘unity’ between the aforesaid two basis
(the normalized mutual coherence is N-2, N being the number of pixels in the image). The mutual
coherence between Fourier and wavelet basis is always higher than unity [31, 32]. Higher mutual
coherence leads to worse signal reconstruction accuracy (when the number of measurement samples is
fixed). In order to alleviate this problem in static MRI, ‘spread spectrum CS’ techniques were proposed in
[31, 32]. In [31] it was reported that for an image of size 256 x 256, the mutual coherence between the
Fourier measurement basis (k-space) and the Dirac basis (pixel domain) is 3.906 × 10−3; and the mutual
coherence between the Fourier measurement basis and the wavelet basis (sparsifying transform) is
2.272×10−2 for Daubechies wavelets with 4 vanishing moments at 3 levels of decomposition. Thus, if one
uses wavelets for sparsifying the difference image, the mutual coherence increases by a factor of 5. Even
though mutual coherence based bounds are pessimistic, it follows that the number of wavelet coefficients
that can be recovered by CS algorithms will be less than number of pixel values recovered. Therefore, the
modeling/reconstruction error for a CS recovery method using wavelet coefficients (such as [17]) will be
more than our proposed method. The slim advantage wavelet transform coefficients have over pixel
values in terms of energy compaction in fewer coefficients (see Fig. 1) is offset by the increased mutual
coherence between the measurement basis and the sparsifying transform.
Following the assumption that the difference image is sparse in the pixel domain, we propose solving for
the difference image via the following problem,

min xt subject to yt  Rt F2 D xt 1  Rt F2 D xt 


p 2
p 2
(25)
xt
There are some fast algorithms to solve the convex problem (i.e. where p=1) such as the SPGL1 [30] and
NESTA [33]. Even though efficient, the algorithms are not fast enough for real-time implementation.
Solvers for the non-convex problem (0<p<1) are either based on the iterative reweighted least squares [7-
9] or on iterative reweighted l1-norm minimization [15, 16]. Both of them are even slower than their
convex counterparts.
In this paper, we are interested in using non-convex lp-norm minimization (25) for reasons discussed in
Section 2. Because there is no off-the-shelf solver that can be employed for this purpose, we derive a fast
algorithm for solving the said problem in sub-section 4.1.
The other advantage of our method over [17] is the reconstruction speed. Both [17] and our proposed
method require solving a CS optimization problem. The optimization problem is solved iteratively. Each
iteration requires a backward and forward projection. For our method, the projection is a Fourier
transform, but for [17], the projection is a wavelet transform followed by a Fourier transform. This
additional application of the wavelet transform requires more time per iteration. In most cases (for MRI
reconstruction), the algorithms for solving the CS optimization problem runs for a specified number of
iterations. Hence, our method is faster, as the computational cost per iteration is lower compared to [17].
Our approach is based on reconstructing the difference frames. In this work, we assume that the first
frame is reconstructed from the fully sampled k-space via inverse Fourier Transform. However, one can
use Compressed Sensing based techniques to reconstruct the initial frame as well from under-sampled k-
space samples.
There is an offline equivalent to our work called the k-t FOCUSS [8, 9]. In essence it solves the following
problem,

min I 2 D  F1D ( x) 1 subject to y  RF2 D x 2  


2

i.e., it assumes the dynamic MRI signal to be sparse in x-f space.


It uses a FOCUSS [26] based method to solve the optimization problem. However in the said studies [8,
9], the standard FOCUSS algorithm has been modified to incorporate additional information regarding
the dynamic MRI sequence. The most updated form of k-t FOCUSS incorporates concepts of motion
estimation and motion compensation from video compression in order to modify the FOCUSS algorithm
[34].
Such concepts can potentially benefit our proposed method as well, i.e. instead of using the previous
frame as the reference image, we can use a more sophisticated reference frame that incorporates motion
estimation. In that situation (assuming that the motion estimation is correct), the difference between the
current frame and the motion estimated and compensated reference frame will be even sparser; this in turn
would yield better recovery results. However, motion estimation and compensation is time consuming as
is well known by researchers in video processing. This would slow down the reconstruction and it would
not be near real-time any more. Speed of reconstruction is not of much importance to offline techniques
like k-t FOCUSS but is essential for us since we aim at near real-time reconstruction.
4.1 Reconstruction Algorithm
Suppose one needs to solve the following constrained optimization problem,
min f ( x) subject to g ( x)   (26)
x

Instead of solving the constrained optimization problem, one can solve its unconstrained Lagrangian
version,
min g ( x)   f ( x) (27)
x

With a correct choice of the parameters τ and λ, the constrained and the unconstrained forms are
equivalent. However, for most practical problems it is not possible to determine λ analytically, given the
value of τ. In such a case, if the Pareto curve between the functions f(x) and g(x) is smooth [29], it is
possible to solve the constrained problem (26) by iteratively solving the unconstrained problem (27) by
cooling λ, i.e. one starts with a high value of λ, solves (27), uses the current solution to initialize the next
problem with a reduced value of λ and continues until a stopping criterion is met (such as g ( x)   ).

In this work, the objective is to solve (25). Comparing (25) with (27) we get f ( x)  xt
p
p
and

g ( x)  yt  Rt F2 D xt 1  Rt F2 Dxt 2 . However solving the constrained problem directly is difficult.


2

Therefore, we would adopt the strategy discussed above. The constrained problem (25) will be solved by
iteratively solving the following unconstrained Lagrangian form with reducing λ.

min yt  Rt F2 D xt 1  Rt F2 D xt   xt


2 p
2 p
(28)
xt

The problem is to derive the solution for (28).


4.1.1 Solution of the Unconstrained Problem
The solution trajectory for the problem defined in (28) is not smooth, especially at the solution point.
Therefore the straightforward gradient based techniques can not be used for minimization. We use the
Majorization-Minimization (MM) framework [35] to solve (28). The MM uses gradient descent
internally. The basic MM algorithm is as follows,
Let J(x) be the (scalar) function to be minimized
1. Set k=0 and initialize x0.
Repeat step 2-4 until suitable a stopping criterion is met.
2. Choose Gk(x) such that
a. Gk ( x)  J ( x) for all x.

b. Gk ( xk )  J ( xk ) .

3. Set xk+1 as the minimizer for Gk(x).


4. Set k=k+1, go to step 2.
For our problem (28) the function to be minimized is re-written as,

J (x)  y  RFx  RF x 2   x
2 p
p

where we have dropped the sub-scripts for ease of notation.


There is no closed form solution to J(x), it must be solved iteratively. At each iteration (k), we choose,

Gk (x)  J (x)  (x  xk )T (aI  F T RT RF )(x  xk )

i.e. Gk (x) || y  RFx  RF x ||22 (x  xk )T (aI  ( RF )T RF )(x  xk )   x
p
p
(29)

In (29), Gk (x) satisfies the condition for MM algorithm when a  1

Now Gk (x) can be alternately expressed as follows,

1
Gk (x)  a || xk  ( RF )T ( y  RFx  RF xk )  x ||22  x p  K
p
(30)
a
where K consists of terms independent of x .
Minimizing (30) is the same as minimizing the following,

Gk (x) || b  x ||22  x
p
p
(31)
a
1 T T
where b  xk  F R ( y  RFx  RF xk ) .
a
Gk (x) is actually decoupled element-wise, i.e.

N2

Gk (x)   (b(i)  x(i)) 2  x(i )
p
(32)
i 1 a

where N2 is the total number of pixels in the image and is the length of the vectors.
Therefore it is possible to minimize (32) element-wise, i.e.
 
Gk (x)  2b(i)  2x(i)  p x(i) p 1 signum(x(i)) (33)
xi a


Setting the partial derivatives Gk (x) to zero and solving gives [35],
xi


xi  signum(b(i)) max(0,| b(i) |  p | b(i) | p 1 ) (34)
2a

This is for each element of the vector. For the entire vector the solution can be compactly represented as,


x  signum(b) max(0,| b |  p diag (| b | p 1 )) (35)
2a

Here diag (| b | p 1 )) represents a diagonal matrix whose elements are | b(i) | p 1 .

This leads to a simple two-step iterative solution for the unconstrained problem (28):
Initialize: x0  0
While k ≤ insweep
1
Step 1. b  xk  F T RT ( y  RFx  RF xk )
a

Step 2. x  signum(b) max(0,| b |  p diag (| b | p 1 ))
2a
End

This is a modified version of the iterative shrinkage algorithm used for solving the unconstrained l1-norm
minimization problem by Iterative Soft Thresholding [35]. This is a first order algorithm as it only
requires a forward and backward Fourier transform in each iteration.
4.1.2 Solution of the Constrained Problem
The goal is to solve the constrained optimization problem (25). As mentioned earlier, the constrained
form will be solved by iteratively solving a series of unconstrained optimization problems (26) with
decreasing values of λ. We have successfully used such cooling techniques is the past for solving non-
convex optimization problems [36-39].
The constrained problem proceeds in two loops. The inner loop solves the unconstrained optimization
problem. The outer loop cools (reduces) the value of λ. There are three parameters to be set – the number
of inner iterations (insweep), the number of outer iterations (outsweep) and the cooling factor (DecFac).
Ideally, one would run the inner loop a large number of times so that the unconstrained problem
converges to the desired solution, and reduce the value of λ slowly (leading to large number of outer
iterations). But such a strategy would lead to a slow algorithm. For us, both the reconstruction accuracy
and the reconstruction time are at a premium. We experimentally found that insweep = 10, outsweep = 2
and DecFac = 0.1 yields good reconstruction results both in terms of speed and accuracy.
The complete algorithm for solving the constrained optimization problem is the following:

Initialize: x0  0 , λ =DecFac*max( F T RT ( y  RFx) )


Outer Loop: While1 l ≤ outsweep
Decrease λ by λ= λ*DecFac
Inner Loop: While2 k ≤ insweep
1
i. b  xk  F T RT ( y  RFx  RF xk )
a

ii. x  signum(b) max(0,| b |  p diag (| b | p 2 ))
2a
iii. Update: k=k+1 and return to step i.
End While2 (inner loop ends)
Update l = l+1
End While1 (outer loop ends)

5. EXPERIMENTAL EVALUATION
The experiments were carried out on a laptop with an AMD (Advanced Micro Devices) 64 bit processor
having 4GB of RAM. The simulations were done in Matlab 2009a environment running on Windows 7.
We have compared our proposed method with four state of the art reconstruction methods – two are
offline methods [5], [6] and two are online [17], [18]. In [5] the spatial and temporal redundancies are
exploited by solving the following Spatio-Temporal Total Variation (TV) minimization problem,

min y  RF2 D x 2  1  (t xi )2  2  (h x)2  (v x) 2


2
(36)
x

where  (t xi )2 accounts for temporal variation and  (h x)2  (v x)2 is the spatial total variation
term.
In spatio-temporal TV, the parameters λ1 and λ2 needs to be specified. Unfortunately these parameters can
not be specified based on rigorous optimization theory. They need to be tuned. In this work, we follow the
tuning mechanism proposed in [5]. The value of λ1 was fixed by using the L-curve method after putting λ2
to zero, thereby using temporal differencing only. Once the value of λ1 was fixed, the value of λ2 was then
chosen by minimizing the error in the reconstruction, as compared with ground-truth. In other words, λ2
needs to be manually tuned.
The other offline technique is called k-t SLR [6]. It exploits the rank deficiency and sparsity of the MR
images in the x-f space in order to reconstruct them. For k-t SLR (18), the parameters λ and ε needs to be
defined. In this work, we are assuming that the data is free from noise, therefore ε=0. For fixing λ, we
follow the manual parameter tuning technique proposed in [6], i.e. results are reported for the particular
value of λ which yields the least error.
It should be noted that in practical situation such manual tuning techniques do not work since the fully
sampled ground-truth are not available. For k-t SLR it is found that the reconstruction is robust to choice
of λ as long as it has a high value.
The online techniques we have compared with are thoroughly discussed in section 4.2. In [18] a Kalman
Filter is used for reconstructing the MR images. This has an efficient implementation and has been
reported [18, 19] to be capable of real-time reconstruction. The other online reconstruction method is the
Modified CS method [30]. The implementation of this algorithm can be found in [40]; however this
implementation is slow and we had to tune it a bit in order to reconstruct the images in a reasonable time.
We propose to reconstruct the images via a non-convex optimization problem (25), we need to specify the
value of p. In our previous work on static MRI reconstruction [14], we found that the best reconstruction
results are obtained with p = 0.8. However for the problem addressed in this work we found that the
reconstruction accuracy is not sensitive to the choice of p between 1 and 0.6. We performed a simple
hypothesis testing (t-test) to find out if the reconstruction accuracies vary significantly with differing
values of p between 0.6 and 0.1 (0.6, 0.7, 0.8, 0.9 and 1). We found that there is no significant difference
in the reconstruction accuracies at 5% confidence interval. We could report results for any value of p in
the said range; however we use p=0.6 throughout this work. The other parameter that we need to specify
in (25) is ε. In this case, we assume that the data is noise-free and our aim is to reconstruct the image as
faithfully as possible to the ground-truth; therefore we use ε=0.
The original Kalman Filter method [18, 19] assumed uniform sampling density over the K-space. In this
work, we use radial sampling which is not uniform (samples the center of the K-space more than the
periphery). Thus the criterion for uniform sampling density is not satisfied in this work. In [18, 19] it is
mentioned that one way to bypass this problem is by re-gridding the K-space samples from the non-
uniform sampling pattern to a uniform sampling pattern. However, such re-gridding is time consuming
and hence not suitable for real-time reconstruction. The focus of this work is on rela-time dynamic MRI
reconstruction. Therefore in this work, we use the Kalman Filter method on the raw samples without re-
gridding. It has been mentioned in [18, 19] that this may decrease the reconstruction accuracy of the
Kalman Filter.
5.1 Datasets
The experimental evaluation was performed on five sets of data. Two myocardial perfusion MRI datasets
were obtained from [41]. These datasets were used in [5]. The two sequences will be called the Cardiac
Perfusion Sequence 1 and 2. The data was collected on a 3T Siemens scanner. In this work we simulated
a radial sampling with 24 lines were acquired for each time frame; this corresponds to an under-sampling
ratio of 0.21. The full resolution of the dynamic MR images is 128 x 128. About 6.7 samples were
collected per second. The scanner parameters for the radial acquisition were TR=2.5–3.0 msec, TE=1.1
msec, flip angle = 12° and slice thickness = 6 mm. The reconstructed pixel size varied between 1.8 mm 2
and 2.5 mm2. Each image was acquired in a ~ 62-msec read-out, with radial field of view (FOV) ranging
from 230 to 320 mm.
The third and fourth datasets comprise of the Larynx and Cardiac sequence respectively. The data has
been obtained from [40]. The larynx sequence is of size 256 x 256 and the cardiac sequence is of size 128
x 128 for each time frame. Six images were collected per second. Unfortunately other details pertaining to
the MRI scan cannot be obtained either from the website [39] or related papers [17, 20, 21]. In this paper,
we have simulated radial sampling with 24 lines for each image; This corresponds to an under-sampling
ratio of 0.11 for the Larynx sequence 0.21 for Cardiac sequence.
Our final dataset is obtained from [42]. It consists of a dynamic MRI scan of a person repeating the word
‘elgar’. This is the Speech Sequence. The image is of resolution 180 x 180 and is obtained at the rate of 6
frames per second. This dataset was collected to study the tongue positions during speech. The MRI
acquisition parameters related to this study are not reported here. For this sequence, we simulated radial k-
space sampling with 24 radial lines which corresponds to an under-sampling ratio of 0.15.
It should be noted that the frame-rates for the Cardiac perfusion sequences are 6.7 frames per second, and
for the Larynx, Cardiac and Speech sequences it is 6 frames per second. The original fully sampled
ground-truth data has been collected at the said rates. The under-sampling ratios used in these
experiments are simulations.
For all the data sequences, radial sampling has been used. This is a non-Cartesian sampling scheme. The
mapping from the Cartesian image space to the Non-Cartesian Fourier space is achieved through the Non-
Uniform FFT [43]. NUFFT was used for our proposed method, as well as Modified CS, k-t SLR and
Spatio-Temporal TV. For Kalman Filtering, the re-gridding operation is handled internally by the
algorithm.
5.2 Results
For all the methods, we assume that the initial frame is obtained from fully sample K-space data via
inverse Fourier transform. Therefore, the reconstructing accuracy and reconstruction time for the first
frame is omitted from the ensuing discussion.
5.2.1 Reconstruction Accuracy
For real-time dynamic MR image reconstruction, we are concerned with two aspects – The accuracy of
reconstruction and the speed of reconstruction. Here we are measuring the reconstruction accuracy in
terms of normalized (also called Relative) mean squared error (NMSE). NMSE (and SNR = -
20log(NMSE)) is a widely used and accepted quantitative evaluation metric for evaluating MR images [5,
17, 35]. However, NMSE does not provide a qualitative measure on the reconstruction accuracy, therefore
we also provide the reconstructed and difference images (difference between reconstructed and ground-
truth). The quantitative error for the five different sequences is shown in the following graphs. In all the
graphs, TV stands for Spatio-Temporal TV and KF stands for Kalman Filter.
Fig. 3. Comparison of reconstruction errors in terms of NMSE (KF applied on raw data without re-gridding on a uniform grid)

The general observation is that the offline techniques yield better reconstruction accuracy than the online
ones. This is expected. Our proposed method is the best online technique in terms of reconstruction
accuracy for all the datasets. The k-t SLR gives better reconstruction results better than the Spatio-
Temporal TV for proper choice of k-t SLR parameters. This too is expected since Spatio-Temporal TV is
a special case of k-t SLR. Of the two online reconstruction techniques we compared against – Modified
CS and Kalman Filter, the former gives better results for the Larynx and Cardiac sequences; for the two
cardiac perfusion sequences and the speech sequence the reconstruction accuracy from both of them are
almost the same.
For the cardiac sequence, it is noted that the reconstruction errors for all the online techniques increase
with time; in other words the error propagates and accumulates. In this work, we did not use any check to
contain such error propagation. However, it is possible to device simple ways to stop this. The simplest
approach is to reconstruct one frame from fully sampled data at periodic intervals (much like I-frames in
video processing). For cardiac sequence, where it is known a priori that the motion will be quasi-periodic,
it is possible to use this information to determine the periods.
For qualitative evaluation we show the ground-truth images, reconstructed images (Fig. 4) and the
difference images (between reconstructed and groundtruth) (Fig. 5). One image each from the cardiac
sequence (frame 5), larynx sequence (frame 17), speech sequence (frame 52) cardiac perfusion sequence
1 (frame 34) and cardiac perfusion sequence 2 (frame 16) is chosen for the purpose (the frame numbers
are randomly chosen). The difference images have been magnified 5 times for better visual clarity.
Fig. 4. Reconstructed Images. Row 1: Ground-truth; Row 2: Proposed Reconstruction; Row 3: Kalman Filter Reconstruction (KF
applied on raw data without re-gridding on a uniform grid) (online); Row 4: Modified CS Reconstruction (online); Row 5:
Spatio-Temporal TV Reconstruction (offline); Row 6: k-t SLR Reconstruction (offline)
Fig. 5. Difference Images. Row 1: Proposed Reconstruction; Row 2: Kalman Filter Reconstruction (KF applied on raw data
without re-gridding on a uniform grid) (online); Row 3: Modified CS Reconstruction (online); Row 4: Spatio-Temporal TV
Reconstruction (offline); Row 5: k-t SLR Reconstruction (offline)

Visual evaluation shows the nature of reconstruction artifacts. This is especially evident from the
difference images. A good reconstruction results in difference images that are mostly dark with no
perceptible structured artifacts. k-t SLR gives the best reconstruction results. The differences images are
almost black. Also the difference images do not show any structural reconstruction artifacts. The Spatio-
Temporal TV method shows little visible reconstruction artifacts. The modified CS method results in
show heavy reconstruction artifacts mostly owing out of radial sampling. The Kalman Filter gives the
worst results. This is especially evident from the cardiac perfusion sequences. The reconstructed images
(Fig. 4) and difference images (Fig. 5) clearly show that, this method is unable to recover certain
structures. The difference image is all white in those areas. Our proposed method shows artifacts, but they
are not structured and are much lower intensity than the Kalman Filter or the Modified CS. Visually,
results from our proposed method seems almost at par with offline techniques and is discernibly better
than the online methods compared against.
For all the experiments, the same radial sampling trajectory was used for all the frames. It does not pose a
problem for CS based online reconstruction techniques like Modified CS and our proposed method. But
using the same sampling pattern for all the frames is not optimal for offline techniques. CS based offline
recovery techniques maximally benefit from incoherent sampling. Ideally, one would use different
sampling patterns for each frame when using offline reconstruction [32]. The following table shows that
the reconstruction error (NMSE for the full sequence) for offline techniques reduce even more when
different radial sampling patterns are used for different frames.
Table 1. Variation in NMSE for differing sampling patterns
Datasets Spatio-Temporal TV k-t SLR

Same Sampling for all Different Sampling for Same Sampling for all Different Sampling for
frames each frame frames each frame

Cardiac Perfusion 1 0.09 0.06 0.07 0.05

Cardiac Perfusion 2 0.11 0.07 0.08 0.05

Larynx 0.03 0.02 0.02 0.02


Cardiac 0.02 0.01 0.02 0.01

Speech 0.26 0.17 0.17 0.11

5.2.2 Reconstruction Speed


As was mentioned earlier, reconstruction accuracy is one part of the dynamic MRI reconstruction
problem. The speed of reconstruction is also at a premium for this case. The following graphs show the
plot of the reconstruction times from our proposed method versus Kalman Filter [18] and Modified CS
[30]. The default parametric settings for the Modified CS method is extremely slow; we tuned the
parameters in order to reduce its reconstruction time by about two orders of magnitude. This tuning does
not affect the reconstruction accuracy perceptibly. The reconstruction time for the offline method is not
given since it reconstructs the whole data sequence in retrospective and not on a frame-by-frame basis.
Fig. 6. Comparison of reconstruction times for different dynamic MRI sequences.

The results show that even though Modified CS [30] is a successful online method, it is not suitable for
real-time reconstruction of images. Even for images of size 128 x 128 (cardiac perfusion sequences and
cardiac sequence), it takes more than a second to reconstruct them. The Kalman Filter yields the fastest
reconstruction as it only requires a fixed number of re-gridding and Fourier transform operations to
reconstruction each time frame. It is about 4 times faster than our proposed method. Our proposed method
is the nearly a magnitude faster than Modified CS – for images of size 128 x 128, the average
reconstruction time per frame is 0.16 seconds, for images of size 180 x 180, the average reconstruction
time is 0.2 seconds and for images of size 256 x 256 the average reconstruction time is 0.4 seconds.
Although this is not perfectly real-time speed, it is near about so. One needs to keep in mind that the
experiments were performed in Matlab. Porting the reconstruction algorithms to a development
environment such as C/C++ will speed up the reconstruction even further.
6. CONCLUSION
In this work, we propose a real time dynamic MRI reconstruction method. To reconstruct a current frame,
the method reconstructs the difference image between the current and its previous image and adds the
difference to the previous image. It is assumed that the difference image is sparse. The motivation for this
work follows from previous studies [4-9], where it is assumed that the difference image between two
consecutive time frames for dynamic MRI sequence is sparse.
We compared our work with state-of-the-art offline and online reconstruction techniques. The offline
reconstruction methods we compared against are Spatio-Temporal Total Variation minimization [5] and
k-t SLR [6]. For online reconstruction we considered the Kalman Filtering method [18] and the Modified
Compressed Sensing technique [30]. We found that our proposed method yields reconstruction results
that are slightly worse than the offline methods but is better than the online methods.
We focus on reconstructing MRI images in real-time. Our method is able to reconstruct 128 x 128 images
at the rate of 6 images per second, 180 x 180 images at the rate of 5 frames per second and 256 x 256
images at the rate of 2.5 images per second. The Kalman Filtering method [18] yields even faster
reconstruction. On an average it is about 4 times faster than our proposed method. However, as seen
before the reconstruction from Kalman Filter is of considerably lower quality. In the future, we need to
port our proposed method from the Matlab simulation environment to C/C++, where even further
performance improvement (reduction in reconstruction time) is expected.
The advantage of our proposed method and Modified CS over Kalman Filtering is that the former ones do
not have a ‘settling time’. The Kalman filter based methods, requires a few time frames to ‘stabilize’ and
yield good results. Our proposed method and Modified CS do not have this requirement, they will yield
good reconstruction results as long as the difference images are sparse.
ACKNOWLEDGEMENT
We are thankful to Dr. Pauly for sharing the Kalman Filtering code with us. This work was supported by
NSERC, the Natural Sciences and Engineering Research Council of Canada and by QNRF, Qatar
National Research Fund No. NPRP 09 – 310 – 1 - 058.

REFERENCES
[1] M. Lustig, D.L Donoho and J.M Pauly “Sparse MRI: The Application of Compressed Sensing for Rapid MR Imaging”
Magnetic Resonance in Medicine, Vol. 58 (6), pp.1182-1195, 2007.
[2] J. Trzasko and A. Manduca, “Highly undersampled magnetic resonance image reconstruction via homotopic l0-
minimization,” IEEE Trans. Med. Imaging, Vol. 28 (1), pp. 106–121, 2009.
[3] U. Gamper, P. Boesiger, S. Kozerke, “Compressed sensing in dynamic MRI”, Magnetic Resonance in Medicine, Vol. 59
(2), pp. 365 – 373, 2008.
[4] L. Chen, M. C. Schabel, E. V. R. DiBella, “Reconstruction of dynamic contrast enhanced magnetic resonance imaging of
the breast with temporal constraints”, Magnetic Resonance Imaging, Vol. 28 (5), pp. 637-645, 2010.
[5] G. Adluru, C. McGann, P. Speier, E. G. Kholmovski, A. Shaaban and E. V. R. DiBella, “Acquisition and Reconstruction of
Undersampled Radial Data for Myocardial Perfusion Magnetic Resonance Imaging”, Journal of Magnetic Resonance
Imaging, Vol. 29, pp. 466-473, 2009.
[6] S. Goud, Y. Hu, E. DiBella, M. Jacob, “Accelerated dynamic MRI exploiting sparsity and low-rank structure”, IEEE Trans.
Medical Imaging, pg 1042-1054, Vol. 30, May 2011.
[7] H. Jung, J. C. Ye, E. Y. Kim, “Improved k-t BLAST and k-t SENSE using FOCUSS”, Phys Med Biol, Vol. 52, pp. 3201–
3226, 2007.
[8] H. Jung, J. Park, J. Yoo, and J. C. Ye, “Radial k-t FOCUSS for High-Resolution Cardiac Cine MRI”, Magnetic Resonance
in Medicine, Vol. 63, pp. 68–78, 2010.
[9] H. Jung, J. Park, J. Yoo, and J. C. Ye, “k-t FOCUSS: A General Compressed Sensing Framework for High Resolution
Dynamic MRI”, Magnetic Resonance in Medicine, Vol. 61, pp. 103–116, 2009.
[10] N. Todd, G. Adluru, A. Payne, E. V. R. DiBella and D. Parker, “Temporally Constrained Reconstruction Applied to MRI
Temperature Data”, Magnetic Resonance in Medicine, Vol. 62, pp. 406–419, 2009.
[11] G. Adluru, S.P. Awate, T. Tasdizen, R.T. Whitaker, E.V.R. DiBella, “Temporally Constrained Reconstruction of Dynamic
Cardiac Perfusion MRI”, Magnetic Resonance in Medicine, Vol. 57, pp. 1027—1036, 2007.
[12] J. Trzasko, A. Manduca and E. Borisch, “Sparse MRI Reconstruction via Multiscale L0-Continuation”, IEEE Workshop on
Statistical Signal Processing, pp. 176-180, 2007
[13] R. Chartrand, “Fast algorithms for nonconvex compressive sensing: MRI reconstruction from very few data”, in IEEE
International Symposium on Biomedical Imaging, 2009.
[14] A. Majumdar and R. K. Ward, “Under-determined Non-Cartesian MR Reconstruction”, MICCAI, pp. 513-520, 2010.
[15] C. H. Chang and J. Ji, “Parallel Compressed Sensing MRI Using Reweighted L1 Minimization”, ISMRM 2011.
[16] E. J. Candès, M. Wakin and S. Boyd, “Enhancing sparsity by reweighted l1 minimization”, J. Fourier Anal. Appl., Vol. 14,
pp. 877-905, 2007.
[17] N. Vaswani, “LS-CS-residual (LS-CS): Compressive Sensing on the Least Squares Residual”, IEEE Trans. Signal
Processing, Vol. 58 (8), pp. 4108-4120, 2010
[18] U. Sümbül, J. M. Santos, and J. M. Pauly, “A Practical Acceleration Algorithm for Real-Time Imaging”, IEEE Transactions
on Medical Imaging, Vol. 28 (12), 2009.
[19] U. Sümbül, J. M. Santos, and John M. Pauly, “Improved Time Series Reconstruction for Dynamic Magnetic Resonance
Imaging”, IEEE Transactions on Medical Imaging, Vol. 28 (7), 2009.
[20] C. Qiu and N. Vaswani, “Compressive Sensing on the Least Squares and Kalman Filtering Residual for Real-time Dynamic
MRI and Video Reconstruction”, Submitted to IEEE Trans. Image Processing in December 2009.
[21] D. Donoho, “Compressed Sensing”, IEEE Trans. on Information Theory, Vol. 52(4), pp. 1289 - 1306, 2006.
[22] E. Candès, J. Romberg, and T. Tao, “Robust uncertainty principles: Exact signal reconstruction from highly incomplete
frequency information”, IEEE Trans. on Information Theory, 52(2) pp. 489 - 509, 2006.
[23] K. A. Khalsa and J. A. Fessler, “Resolution properties in regularized dynamic MRI reconstruction”, IEEE International
Symposium on Biomedical Imaging, pp. 456-459, 2007.
[24] M. Lustig, J.M. Santos, D.L. Donoho, J.M Pauly “k-t SPARSE: High Frame Rate Dynamic MRI Exploiting Spatio-
Temporal Sparsity” ISMRM '06.
[25] B. D. Rao and K. Kreutz-Delgado, “An affine scaling methodology for best basis selection,” IEEE Transactions on Signal
Processing, Vol.47 (1), pp.187-200, 1999.
[26] B. Zhao, J. P. Haldar, C. Brinegar, Z. P. Liang, “Low rank matrix recovery for real-time cardiac MRI”, International
Symposium on Biomedical Imaging, pp.996~999, 2010.
[27] B. Recht B, W. Xu, B. Hassibi, “Conditions and Thresholds for Rank Minimization”, Mathematical Programming. Ser B.
Vol. 127, pp. 175-211. 2011
[28] B. Recht, M. Fazel, and P. A. Parrilo, “Guaranteed Minimum Rank Solutions to Linear Matrix Equations via Nuclear Norm
Minimization”, SIAM Review. Vol 52 (3), pp. 471-501, 2010.
[29] E. van den Berg and M. P. Friedlander, “Probing the Pareto frontier for basis pursuit solutions”, SIAM J. on Scientific
Computing, Vol. 31(2), pp. 890-912, 2008.
[30] N. Vaswani and W. Lu, “Modified-CS: Modifying Compressive Sensing for Problems With Partially Known Support”,
IEEE Transactions on Signal Processing, Vol.58 (9), pp.4595-4607, 2011.
[31] Y. Wiaux, G. Puy, R. Gruetter, J. P. Thiran, D. Van De Ville, D and P. Vandergheynst, “Spread spectrum for compressed
sensing techniques in magnetic resonance imaging”, IEEE International Symposium on Biomedical Imaging, pp. 756 – 759,
2010.
[32] G. Puy, P. Vandergheynst, R. Gribonval and Y. Wiaux, “Universal and efficient compressed sensing by spread spectrum and
application to realistic Fourier imaging techniques”, EURASIP Journal on Advances in Signal Processing, Vol. 6, 2012.
[33] S. Becker, J. Bobin, and E. J. Candès “NESTA: a fast and accurate first-order method for sparse recovery”, SIAM J. on
Imaging Sciences, Vol. 4(1), 1-39, pp. 2009.
[34] H. Jung and J. Ye, “Motion estimated and compensated compressed sensing dynamic magnetic resonance imaging: What we
can learn from video compression techniques,” International Journal of Imaging Systems and Technology, vol. 20, no. 2, pp.
81–98, 2010.
[35] http://cnx.org/content/m32168/latest/
[36] A. Majumdar and R. K. Ward, “Some Empirical Advances in Matrix Completion”, Signal Processing, Vol. 91 (5), pp. 1334-
1338, 2011.
[37] A. Majumdar and R. K. Ward, “An Algorithm for Sparse MRI Reconstruction by Schatten p-norm Minimization”, Magnetic
Resonance Imaging, Vol. 29(3), pp. 408-17, 2011.
[38] A. Majumdar and R. K. Ward, “Exploiting Rank Deficiency and Transform Domain Sparsity for MR Image
Reconstruction”, Magnetic Resonance Imaging (accepted).
[39] A. Majumdar and R. K. Ward, “Compressive Color Imaging with Group Sparsity on Analysis Prior”, IEEE International
Conference on Image Processing, pp. 1337-1340, 2010.
[40] http://home.engineering.iastate.edu/~luwei/modcs/
[41] http://www.sci.utah.edu/bisti.html
[42] http://www.phon.ox.ac.uk/jcoleman/Dynamic_MRI.html
[43] J.A. Fessler, “On NUFFT-based gridding for non-Cartesian MRI”, Journal of Magnetic Resonance, Vol. 188(2), pp. 191–
195, 2007.

You might also like