An Analytical Comparison of Different Regularization Parameter Selection Methods for Regularization in Auto-Calibrating Parallel MRI

An Analytical Comparison of Different Regularization
Parameter Selection Methods for Regularization in
Auto-Calibrating Parallel MRI
Submitted by
Saradindu Sengupta
Reg. No: 91616006
In partial fulfilment of the requirements for the award of
Master of Science in Computer Science with Specialization in Machine
Intelligence
1 | P a g e

April 2017
BONAFIDE CERTIFICATE
This is to certify that the mini project entitled “An
Analytical Comparison of Different Regularization
Parameter Selection Methods for Regularization in
Auto-Calibrating Parallel MRI” submitted by SARADINDU
SENGUPTA in partial fulfillment of the requirement for the award of the Master
of Science in Computer Science with specialization in Machine Intelligence
is a bonafide record of the work under my guidance and supervision at IIITM-K.
Signature of the Head of the Institute
Dr. Rajasree M. S.
Director & Professor, IIITM-K
2 | P a g e
Signature of the Supervisor Signature of the Course Coordinator
Dr. Joseph S. Paul Dr. Asharaf S.
Professor, IIITM-K Professor, IIITM-K

DECLARATION
I hereby declare that the mini project entitled “An Analytical
Comparison of Different Regularization Parameter Selection Methods for Regularization in
Auto-Calibrating Parallel MRI”, submitted in partial fulfillment of the requirements for the ward
of degree of Master of Science in Computer Science with specialization in Machine
Intelligence is substantially the result of my own work except where explicitly indicated in the text
carried out during the second semester under the guidance of Dr. Joseph Paul, Associate Professor,
IIITM-K. The report may be freely copied and distributed provided the source is acknowledged.
Date : 01.05.2017 Saradindu Sengupta
Place: Thiruvananthapuram Reg. No: 91616006
3 | P a g e

ACKNOWLEDGEMENT
The culmination of my project requires more than one word of thanks to all those who were a part
of these efforts that I had undertook for this projection and also in preparing the report.
I am extremely grateful to Dr. Rajasree M S, for providing me with good facilities and a proper
environment which helped me in enhancing my ability for undertaking a project of this scale. I
thank Dr. Joseph S. Paul, Associate Professor, IIITM-K, Thiruvananthapuram, my guide for his
valuable suggestion, appraisal and guidance. I express my heartfelt thanks to Dr. Asharaf S,
Assistant Professor, IIITM-K, Thiruvananthapuram, for his valuable support towards the successful
completion of my project. I also thank all the teaching and non-teaching staff for their co-operation
and support throughout.
I also thank my seniors and all our friends for helping me during various stages of my work. I truly
admire my parents for their constant encouragement and enduring support which was inevitable for
my success of my ventures and completion of this work.
Saradindu Sengupta
4 | P a g e

Content Page No.
Abstract 6
1.Introduction 7
2.Scope of the Study 9
3.Requirement
Specification
3.1.Data sets 10
3.2.Technical Requirement 11
4.Methods &
Materials
4.1Methods 4.1.1. GRAPPA 12
4.1.2. SPIRiT 16
4.2Materials 4.2.1. Regularization 20
4.2.2. Optimum parameter selection Strategies for
regularization
21
5.Results 5.1GRAPPA Results 21
5.2SPIRiT Results 22
6.Discussion 23
7.Conclusion 23
Literature Survey 24
List of symbols 25
List of Figures & Tables 26
References 28
5 | P a g e

Abstract
This study analyzes the optimality of regularized reconstruction in auto calibrating parallel MRI
GRAPPA and SPIRiT. Regularization leads to acceleration in parallel imaging and the selection of
regularization factors leads to an optimal reconstruction of the acquired data. This study
analytically studies various strategies for obtaining the regularization parameter using L-curve and
GCV for spectral cut-off regularization in SPIRiT and GRAPPA image reconstruction for optimal
regularization. Performance variations due to adaptation of different methodologies are evaluated.
In this study only spectral cut-off regularization has been studied using generalized cross validation
and L-curve as regularization parameter selector. All the experiments have been performed on
seven set of volunteer data sets procured through Siemens 1.5T Magnetom-Avanto clinical MR
scanner. The experiment results show that presence of sharp L-corner does not guarantee in artifact
free image; For same no of ACS lines and sampling factor sharp L-corners are better observed in
SPIRiT compared to GRAPPA reconstruction; parameter values based on GCV always exceeds of
that L-curve. Minimization in noise propagation during reconstruction phase and diminishing the
effect of aliasing can be achieved with the effective use of regularized solution.
Keywords: GRAPPA, SPIRiT, image-reconstruction, parallel-MRI, L-curve, GCV, Regularization-
factor, spectral cut-off.
6 | P a g e

Introduction
Parallel MRI is a robust process of speedy acquisition of MRI data. Parallel imaging is a widely-
used technique where the known placement and sensitivities of receiver coils are used to assist
spatial localization of the MR signal. Having this additional information about the coils allows
reduction in number of phase-encoding steps during image acquisition. This, in turn, potentially
results in a several-fold reduction in imaging time. In parallel MR imaging, a reduced data set in
the phase encoding direction(s) of k-space is acquired to shorten acquisition time, combining the
signal of several coil arrays. The spatial information related to the phased array coil elements is
utilized for reducing the amount of conventional Fourier encoding.
First, low-resolution, fully Fourier-encoded reference images are required for sensitivity
assessment. Parallel imaging reconstruction in the Cartesian case is efficiently performed by
creating one aliased image for each array element using discrete Fourier transformation. The next
step then is to create a full FOV (field of view) image from the set of intermediate images. Parallel
reconstruction techniques can be used to improve the image quality with increased signal to noise
ratio, spatial resolution, reduced artifacts, and the temporal resolution in dynamic MRI scans.
Parallel imaging algorithms can be divided into 2 main groups:
 Image reconstruction produced by each coil (reconstruction in the image domain, after
Fourier transform): SENSE (Sensitivity Encoding), PILS (Partially Parallel Imaging with
Localized Sensitivity), ASSET.
 Reconstruction of the Fourier plane of images from the frequency signals of each coil
(reconstruction in the frequency domain, before Fourier transform): GRAPPA, SPIRiT.
Additional techniques include SMASH, SPEEDER and mSENSE (an image based enhanced
version of SENSE).
The second type of technique is referred as “data driven reconstruction” because they are based on
limited knowledge of the underlying physical process and rely on training data to calibrate the
relationship between input and output data. Both types of reconstruction require calibration for
either coil sensitivity estimation or interpolation coefficients for k-space reconstruction.
7 | P a g e

 Non-Parallel MRI vs Parallel MRI
For example, a set of 4 receiver coils surrounding a head being used to detect the MR signal arising
from a pixel. Because each coil is at a different distance from the pixel, the signal recorded by each
coil varies as a function of position (closer coils have stronger signals). Just by looking at the
relative intensities from each coil, it is possible to make a crude prediction of the approximate
source of the MR signal.
In "regular" (non-parallel) imaging, multiple surface coils may be used to detect the MR signal, but
their individual outputs are combined into one aggregate complex signal that is digitized and
processed into the final image. In parallel imaging, conversely, the signals from individual coils are
amplified, digitized, and processed simultaneously "in parallel" along separate channels, retaining
their identities until near the end.
8 | P a g e

Scope of the study
In our study of different regularization methods, both types of reconstruction techniques need
calibration to estimate coil sensitivity or interpolation coefficients for k-space reconstruction. It is
performed by finding the solution of a system of equations connecting matrix and the
corresponding set of observations from each coil.
The two most popular regularized methods are
 Spectral cut off; also, known as Truncated SVD(TSVD)
 Tikhonov Regularization; also, known as weight decay, ridge regression or Wiener filtering
In TSVD method the expansion of the SVD solution in truncated where in Tikhonov regularization
an extra damping is added to each SVD component of the solution. As performance of the
regularization methods depend on the choice of regularization parameter (λ) in both cases in
Tikhonov and TSVD methods, a quantitative analysis of the reconstruction performance is required
in order to analyze the appropriateness of the standard selection properties for both GRAPPA and
SPIRiT reconstruction algorithms.
In this study we, analyze the different strategies for defining optimum regularization parameter in
auto-calibrating parallel MRI using both GRAPPA and SPIRiT image reconstruction. Variations
arising from various strategies used determining the regularization parameter are evaluated using
reconstruction error on basis of number of iterations needed. The methods adopted for
determination of regularization parameter are L-curve and GCV. In both cases (i.e. SPIRiT and
GRAPPA) both parameter selection techniques are used at under-sampling rate (R) for 2 and 3.
In this study, only one regularization method is quantitatively analyzed, i.e. Truncated SVD, using
L-curve and GCV at different under-sampling rate (2,3). The reconstruction error analyzed from
the results are then used to determine a viable strategy to choose an optimum regularization
parameter.
9 | P a g e

Requirements Specification
 Data sets:
Raw k-space data were acquired from Siemens 1.5T Magnetom-Avanto clinical MR scanner at
Sree Chitra Tirunal Institute of Medical Science and Technology, Trivandrum, India. All subjects
were volunteers with prior written informed consent collected before scanning. The data consists of
fully sampled brain images acquired using array head coil and spine images acquired using surface
coils. The data sets used in this study are:
Dataset—I
16-channel brain data acquired using FLAIR sequence (TE = 89ms, TR = 9000ms, Tl: 2500ms, in
—plane resolution = 1.0 mm, slice thickness = 5.0 mm FOV = 240 mm).
Dataset-II
6-channel brain data using T2-weighted spin echo sequence (TE = 92 ms, TR = 4500 ms, in-plane
resolution = 1.0 mm, slice thickness = 5.0 mm, FOV = 240 mm).
Dataset—II
32-channel brain data SW1 gradient echo sequence (TE=15ms, TR=349 ms, in- plane resolution =
1.0 mm, slice thickness = 2 mm, F 0V == 230 mm, Hip angle=200).
Dataset-IV
12-channe1 physical phantom T2 weighted spin echo sequence (TE =10 ms, TR = 200 ms, in-plane
Dataset-V
14-channel spine data acquired using T2 weighted spin echo sequence (TE = 98 ms, TR = 4500
ms, in-plane resolution = 1.0 mm, slice thickness = 5.0 mm, FOV = 400 mm).
Dataset-VI
4-channe1 spine data using T2-Weighted spin echo sequence (TE = 98 ms, TR = 3520 ms, in-plane
Dataset-VII
10 | P a g e

6-channel brain data using FLAIR spin echo sequence (TE = 891113, TR = 9000ms, TI= 2500ms,
in-plane resolution=1.0 mm, slice thickness = 5.0 mm, FOV == 240 mm).
 Technical Specification:
The experiments were conducted in MATLAB (version 2015a) using openly available templates
for experiments in GRAPPA and SPIRiT. The figures (curves and reconstructed images then were
saved in MATLAB figure format).
Methods and Materials
 Methods:
 Generalized Auto-calibrating Partially Parallel Acquisitions(GRAPPA)
An auto-calibration scan (ACS) can help to improve the reconstructions in k-space based parallel
MRI, such as GRAPPA. GRAPPA or Generalized Auto-calibrating Partially Parallel Acquisitions is
a parallel imaging technique for fast image acquisition. The frequency domain of image is
reconstructed from the frequency signal of each coil. This information is obtained from several k-
space lines which are acquired in addition to the normal image acquisition. In self-calibrating
parallel magnetic resonance imaging, variable density sampling in k-space is typically employed
along the phase encoding(PE) direction, wherein low spatial frequency signals are fully sampled
while high spatial frequency signals are under-sampled by an outer reduction factor. Usually ACS
is acquired at the center of k-space for higher SNR, thus constituting of full-FOV low-resolution
prior images in the channels of an RF coil array Missing signals are reconstructed directly in k-
space using the spatial correlation, which is already calculated in the central region of k-space, with
their neighboring measured signals. The spatial correlation, also known as the convolution kernel is
calculated during calibration, in which multiple measured signals from all coils close to a target
signal, and then fit to the corresponding signals in an individual coil k-space.
11 | P a g e

Fig a: - GRAPPA work-flow diagram
Calibration is performed using samples from ACS lines, forming the training data. For acquisition
with number of coils = nc , the weights are dependent on an index r=1, 2, R-1, that reflects the
distance of an unacquired from its nearest acquired line. Considering ky to be the PE index of
12 | P a g e

an acquired line, the basic GRAPPA model can be described as one trying to fit the target data
K
1
(ky +r ∆k y ,kx ) to the nearest source data at location ( kx ,k y ). Hence we can represent this
model using the equation
K
1
(ky +r ∆k y ,kx )=zr
1
k(kx ,ky)
Where
kx ,k y
k ¿
). represents a collection of samples. Let us assume the kernel size is [a×b] ,
where a is the number of lines and b is the number of samples of each line included in the kernel.
At first, the calibration weights zr
l
are obtained by inversion of the equation for all training pairs
{k (ky ,kx ), K
1
(ky +r Δk y ,kx )} within ACS. ky refers to only those lines within the ACS that
follow the ordering of acquired lines in the under sampled k region. Each point in the ACS lines
contribute to a single row
kx ,k y
k ¿
). in the calibration matrix M. The observation vector ku for
calibration is obtained from the formula:
zr
l
=(M
H
M)
−1
M
H
ku
Following calibration, estimation of a missing k-space value is achieved by applying the filter
zr
l
to the nearest acquired dataset k(ky,kx). On generalizing the expression for GRAPPA
reconstruction, the missing k-space lines in the j’th
coil is obtained as
K
j
(k y+r ∆ky ,kx)=∑
l=1
nc
∑
b=−Pl
Ph
∑
h=−Fl
Fh
K
l
(ky +bR ∆ky ,kx +h∆ kx )zr
l
Here, b and h are sampling indices of the neighboring points along phase and encoding directions
respectively.
Once all of the lines are reconstructed for a particular coil, a Fourier transform can be used to
generate the uncombined image for that coil. Once this process is repeated for each coil of the
13 | P a g e

Fig b: - Signal acquisition from different coil
array, the full set of uncombined images can be obtained, which can then be combined using a
normal sum of squares reconstruction. By increasing size of the kernel, GRAPPA includes more
information into the estimation, thus improving its fit.
Two main categories of error exist with the GRAPPA technique: model error and noise-related
error. Model error has two components: one from using a limited number (as well as position) of
ACS lines instead of the true coil sensitivity maps and the other from using a limited kernel size.
Noise-related error arises from noise in the measured data and includes noise-induced errors that
occur during kernel weights estimation, mainly due to the matrix inversion process, and errors that
result from the application of the weights to noisy measured data. error. As with any fitting
approach, the model error is expected to decrease with increasing kernel size, while the noise-
related error is expected to increase with the kernel size. To date, the problem of how to choose a
kernel support that optimizes the tradeoff between these errors has not been fully addressed. The
choice of the kernel support has been shown to depend on the coil configuration, noise level in the
acquired data, imaging field of view and orientation, and number and position of ACS lines.
Therefore, GRAPPA implementations employing a fixed kernel support for all situations, as
commonly used, are unlikely to be optimal.
 Iterative Self-Consistent Parallel Imaging Reconstruction (SPIRiT):
14 | P a g e

Fig c: - SPIRiT work-flow diagram
The key approach is to separate the consistency constraints into two:
a) Consistency with calibration
b) Consistency with data acquisition
 Consistency with calibration
Let xi (r) denote an unacquired k-space value at position r,in the ith coil, Rir denote a vector
of all points in the neighborhood of xi (r) on a Cartesian grid and Rir
p
denote a subset of
Rir containing only acquired points. Consistency is enforced in the grid, i.e., xi (r), and its
entire neighborhood across all coils. It is important to emphasize that the notion, entire
neighborhood, includes all the k-space points near xi(r) in all coils whether they were acquired or
not.
Given this the consistency equation for all k-space positions is given by,
Xi (r)=∑
j
(gji, Rir )
Here gji is a full kernel independent of the actual k-space sampling pattern and is the same for
all k-space positions.
The calibration kernel can be written as
15 | P a g e

X=Gx
The matrix G is in fact a series of convolution operators that convolve the entire k-space with the
appropriate calibration kernels,
Xi=∑
j
(gji, Xj)
Applying the operation G on x is the same as attempting to synthesize every point from its
neighborhood. If x is indeed the correct solution, then synthesizing every point from its
neighborhood should yield the exact same k-space data.
 Consistency with data acquisition
This constraint can also be expressed as a set of linear equations in matrix form. Let y be the vector
of acquired data from all coils (concatenated). Let the operator D be a linear operator that relates
the reconstructed Cartesian k-space, x, to the acquired data. The data acquisition consistency is
given by
Y=Dx
This formulation is very general in the sense that x is always Cartesian k-space data, whereas y can
be data acquired with arbitrary k-space sampling patterns. In Cartesian acquisitions, the operator D
selects only acquired k-space locations. The selection can be arbitrary, uniform, variable density or
pseudo random patterns. In non-Cartesian sampling, the operator D represents an interpolation
matrix. It interpolates data from a Cartesian k-space grid onto non-Cartesian k-space locations in
which the data were acquired.
16 | P a g e

The steps for non-Cartesian k-space based reconstruction are as follows:
Fig d: - SPIRiT Reconstruction stages
17 | P a g e

 Materials:
 Regularization:
Ill-conditioning causes noise amplification during image reconstruction, which initially limited the
application of parallel imaging to only moderate acceleration. This limitation can be overcome by
incorporating prior knowledge about the image using regularization methods. In the simplest case,
regularization may consist of a basic quadratic penalty in the framework of a linear reconstruction,
or make use of much more sophisticated techniques which exploit the structure of images but
demand a non-linear reconstruction.
Regularization can be interpreted as prior knowledge and the optimizer as a maximum a posteriori
(MAP) estimate of the image. Although regularization leads to a fundamental trade-off between
bias and noise - which has to be chosen carefully for optimal image quality - it makes the use of
higher acceleration possible. An optimal estimate in terms of mean squared error can only be
obtained with regularization. For optimal results, the prior knowledge should include as much
specific knowledge about the image as possible. For example, regularization can exploit
smoothness in the time domain. Although regularization leads to a fundamental trade-off between
bias and noise - which has to be chosen carefully for optimal image quality - it makes the use of
higher acceleration possible. An optimal estimate in terms of mean squared error can only be
obtained with regularization.
 Optimum parameter selection Strategies for regularization:
A regularized estimate z is a tradeoff between regularization(bias) and perturbation error(variance).
With increase in regularization the regularization error increases and perturbation error decreases.
The optimum parameter selection strategies aim to minimize the sum of bias and variances.
Current technique for regularization parameter selection are discrepancy principle, generalized
cross-validation and L-curve.
In this study regularized versions of GRAPPA and SPIRiT reconstructions using spectral cutoff
regularization (Truncated SVD) are compared using different parameter selection strategies:
A) L-curve
B) GCV
A. L-curve method for selection of truncation parameter in Truncated SVD
18 | P a g e

In TSVD regularization, the L-curve is plotted as a log-log function of residual norm
‖∏ zk−Ku‖ versus the solution norm ‖Lzk‖ for varying parameter k. In TSVD the filter
factor is 1 if he ith
singular value is greater than or equal to the kth
singular value and 0 otherwise.
Thus TSVD corresponds to a spectral cut-off filter that simply cuts off the last (n-k) components.
B. GCV method for selection of truncation parameter in Truncated SVD
In TSVD, the regularization parameter K is chosen so as to minimize the GCV function
G (K )=
‖∏Zk−Ku‖2
2
T
2
Where T=n-K which is effective number of degrees of freedom
19 | P a g e

RESULTS
 GRAPPA results:
All reconstructions are performed on previously acquired data with acceleration factor 2 and 3
respectively with varying ACS lines depending on the data set [Table-1] by computing optimum
kernel size.
Table 1 Here
 L-curve method for selecting truncation parameter in TSVD:
Fig-1 and Fig-5 shows the plots of solution norm versus residual norm for data sets I to VII. The
circles indicate values of regularization parameter at each point on the L-curve. The optimum
parameter is at the corner of the L-curve. Dataset II and III shows sharp L-curve corner. Images
reconstructed using parameter chosen from L-curve corner are shown in Fig-2 and Fig-6.
 GCV method for selecting truncation parameter in TSVD:
Fig-3 and Fig-7 shows the GCV values plotted as function of the truncation parameter using TSVD
regularization. The k -values are indicated by circles and the optimum values are indicated as
asterisks (*). The reconstructed images are shown in Fig-4 and Fig-8. The optimum parameter
values are higher than in L-curve in some cases.
 SPIRiT results:
 L-curve method for selecting truncation parameter in TSVD:
Calibration matrix is constructed using both acquired and non-acquired samples within ACS. The
resulting L-curves are shown in Fig-9 and Fig-13. The circles indicate values of regularization
parameter at each point on the L-curve. Image reconstructed using the parameter values are shown
in Fig-10 and Fig-14. The residual aliasing in Data Set V and VI is evident. Data sets II, III, VII
shows sharp l-curve. All other data sets show similar performance as GRAPPA.
 GCV method for selecting truncation parameter in TSVD:
20 | P a g e

Fig-11 and Fig-15 shows the GCV values plotted as a function of the truncation parameter. The
optimum parameter is indicated using asterisk (*). It is evidently clear that optimum parameter
values are higher than L-curve. The reconstructed images are show in Fig-12 and Fig-16. Data set
V shows significant amount of aliasing and data set VI shows quiet amount of noise.
Discussions
Selection of appropriate regularization parameter leads to accuracy in regularized outputs. In
Parallel MRI, the optimum parameter controls the degree of regularization and thus the
compromise in SNR and artifacts. Under-regularization leads to residual noise or artifacts in the
image, over regularization removes image features and enhances aliasing. In this study, L-curve
and GCV are adopted for optimum parameter selection in spectral cut-off regularization for SPIRiT
and GRAPPA.
Only in a subset of the data sets L-curve show sharp corner though an existence of sharp L-corner
does not guarantee in artifacts free image construction. Although artifacts always appear where
sharp L-curves are not present. With increasing matrix size the L-curve stops behaving
consistently leading to over regularization.
Also, GCV shows minimum only in a handful of data sets where in most of the sets GCV is flat
with multiple local minima.
Conclusion
As both the reconstruction techniques involves discrete ill-posed least squares noisy reconstruction
and artifacts are expected. Though proper and efficient use of regularization and optimal parameter
selection can lead to a healthy reconstruction. But the choice of regularization parameter is only
reasonable in the presence of a knowledgeable error bound. Without such information a priori,
regularization leads to sub-optimal results.
Reconstruction error is much more less in TSVD based on GCV merit function. In all cases L-
curve based regularization parameter values are higher than those in GCV though sharp L-curve
are not evident throughout the data sets and presence of sharp L-corner does not guarantee in
artifact free reconstruction.
For future working scope, the analytically study of Tikhonov regularization for L-curve and GCV
based on this study can be a feasible comparison.
21 | P a g e

Literature Survey
 Tikhonov Regularization:
Tikhonov regularization is the most commonly used method of regularization of ill-posed
problems. Suppose that for a known matrix A and vector b, we wish to find a vector x such that
Ax=b
The standard approach is ordinary least squares linear regression. However, if no x satisfies the
equation or more than one} x does — that is, the solution is not unique — the problem is said to be
ill posed. In such cases, ordinary least squares estimation leads to an overdetermined (over-fitted),
or more often an underdetermined (under-fitted) system of equations. Most real-world phenomena
have the effect of low-pass filters in the forward direction where A maps x to b. Therefore, in
solving the inverse-problem, the inverse mapping operates as a high-pass filter that has the
undesirable tendency of amplifying noise (eigenvalues / singular values are largest in the reverse
mapping where they were smallest in the forward mapping).
 L-Curve:
The L-curve is a log-log plot of the norm of a regularized solution versus the norm of the
corresponding residual norm. It is a convenient graphical tool for displaying the trade-off between
the size of a regularized solution and its fit to the given data, as the regularization parameter varies.
The L-curve thus gives insight into the regularizing properties of the underlying regularization
method, and it is an aid in choosing an appropriate regularization parameter for the given data. In
this chapter, we summarize the main properties of the L-curve, and demonstrate by examples its
usefulness and its limitations both as an analysis tool and as a method for choosing the
regularization parameter
 Generalized Cross Validation:
Cross-validation, sometimes called rotation estimation, is a model validation technique for
assessing how the results of a statistical analysis will generalize to an independent data set.
 ASSET & SENSE:
SENSE (SENSitivity Encoding) and ASSET (Array coil Spatial Sensitivity Encoding) are among
the most widely used parallel imaging methods. These techniques are primarily performed in image
space after reconstruction of data from the individual coils. (This contrasts with GRAPPA/ARC
methods which operate primarily on k-space data before image reconstruction). Each major MR
22 | P a g e

vendor offers some version of the SENSE technique under different trade names: Siemens
(mSENSE), GE (ASSET), Philips (SENSE), Hitachi (RAPID - "Rapid Acquisition through
Parallel Imaging Design"), and Toshiba (SPEEDER).
 ACS:
An auto-calibration scan (ACS) can help to improve the reconstructions in k-space based parallel
MRI, such as GRAPPA. Usually ACS is acquired at the center of k-space for higher SNR, thus
constituting of full-FOV low-resolution prior images in the channels of an RF coil array.
List of Abbreviations
1. pMRI - Parallel Medical Resonance Imaging
2. SENSE - SENSitivity Encoding
3. ASSET - Array Coil Spatial Sensitivity Encoding
4. GRAPPA - Generalized Auto Calibrating Partially Parallel Acquisitions
5. PILS -Partially Parallel Imaging with Localized Sensitivity
6. SPIRiT - Iterative Self-Consistent Parallel Imaging Reconstruction
7. ACS - Auto Calibration Signal / Auto Calibrating Scan
8. FOV - Field of View
23 | P a g e

List of Figures
Fig 1 -: - L-Curve for TSVD regularization using GRAPPA with R=2.
Fig. 2: - GRAPPA reconstructed images using TSVD regularization with L-Curve based
parameter selection
Fig. 3: - GCV for TSVD regularization using GRAPPA with R=2.
Fig. 4: - GRAPPA reconstructed images using TSVD regularization with GCV based
parameter selection.
Fig. 5: -L-Curve for TSVD regularization using GRAPPA with R=3.
Fig 6: - GRAPPA reconstructed images using TSVD regularization with L-Curve based
Fig. 7: - GCV for TSVD regularization using GRAPPA with R=3.
Fig 8: - GRAPPA reconstructed images using TSVD regularization with GCV based
Fig. 9: - L-Curve for TSVD regularization in SPIRiT with R =2
Fig. 10: - SPIRiT reconstructed images using TSVD regularization with L-curve based
parameter selection
parameter selection
Fig. 11: - GCV for TSVD regularization in SPIRiT with R =2
Fig 12: - SPIRiT reconstructed images using TSVD regularization with GCV based
parameter selection
Fig. 13: - L-Curve for TSVD regularization in SPIRiT with R =3
parameter selection
Fig. 15: - GCV for TSVD regularization in SPIRiT with R =3
Fig 16: - SPIRiT reconstructed images using TSVD regularization with GCV based
parameter selection
24 | P a g e

List of Tables
Table 1: - Description of datasets I to VII
Table 2: - Reconstruction error of GRAPPA reconstructed images using TSVD regularization
method
Table 3: - Reconstruction error of SPIRiT reconstructed images using TSVD regularization
method
Table 4: - Parameters of cross-over determination in GRAPPA calibration for datasets I to
VII with R=2
Table 3: - Parameters of cross-over determination in GRAPPA calibration for datasets I to
VII with R=3
Table 6: - Parameters of cross-over determination in SPIRiT calibration for datasets I to VII
with R=2
Table 7: - Parameters of cross-over determination in SPIRiT calibration for datasets I to VII
with R=3
25 | P a g e

Reference
[1] Michael Lustig, John M. Pauly. SPIRiT: Iterative Self-Consistent Parallel Imaging
Reconstruction from Arbitrary k-Space. Magn Reson Med. 2010 August; 64(2): 457–471.
doi:10.1002/mrm.22428.
[2] W. Scott Hoge, Dana H. Brooks. Using GRAPPA to improve auto-calibrated coil sensitivity
estimation for the SENSE family of parallel imaging reconstruction algorithms.
[3] Magn Reson Med. 2008 August; 60(2): 462–467. doi:10.1002/mrm.21634.
[4] Arjun Arunachalam, Alexey Samsonov, Walter F. Block. Self-Calibrated GRAPPA Method for
2D and 3D Radial Data. Magnetic resonance in Medicine (2007).
[5] Anagha Deshmane, MEng, Vikas Gulani, MD, PhD, Mark A. Griswold, PhD, Nicole
Seiberlich, PhD. Parallel MR Imaging. J Magn Reson Imaging. 2012 July; 36(1): 55–72.
doi:10.1002/jmri.23639.
[6] D. S. Weller, J. R. Polimeni2,3, L. Grady, L. L. Wald2,3, E. Adalsteinsson, V. Goyal. SpRING:
Sparse Reconstruction of Images using the Null space method and GRAPPA. Proc. Intl. Soc.
Mag. Reson. Med. 19 (2011).
[7] Leslie Ying, Dan Xu, Zhi-Pei Liang. ON TIKHONOV REGULARIZATION FOR IMAGE
RECONSTRUCTION IN PARALLEL MRI. Proceeding of the 26th
Annual International
Conference, IEEE EMBS, 2007.
[8] Roger Nana, Tiejun Zhao, Keith Heberlein, Stephen M. LaConte, and Xiaoping Hu. Cross-
Validation-Based Kernel Support Selection for Improved GRAPPA Reconstruction. Magnetic
resonance in Medicine (2008).
26 | P a g e

[9] Peng Qu, PhD, Chunsheng Wang, PhD, and Gary X. Shen, PhD. Discrepancy-Based Adaptive
Regularization for GRAPPA Reconstruction. JOURNAL OF MAGNETIC RESONANCE
IMAGING 24:248 –255 (2006).
[10]Daniel S. Weller1, Sathish Ramani, Jon-Fredrik Nielsen, and Jeffrey A. Fessler. Monte Carlo
SURE-Based Parameter Selection for Parallel Magnetic Resonance Imaging Reconstruction.
Magn Reson Med. 2014 May; 71(5): 1760–1770. doi:10.1002/mrm.24840.
[11] Martin Uecker . Parallel Magnetic Resonance Imaging. July2007.
[12] Chunli Wu,Wenjuan Hu,Ruwen Kan. An improved GRAPPA image reconstruction
algorithm for parallel MRI. Control and Decision Conference (CCDC), 2011 Chinese,
IEEE Xplore: 01 August 2011.
[13]Stefan O. Schönberg, Olaf Dietrich, Maximilian F Reiser. Parallel Imaging in Clinical MR
Applications. Springer Science & Business Media, 11-Jan-2007
[14] University of California, Berkley
27 | P a g e

An Analytical Comparison of Different Regularization Parameter Selection Methods for Regularization in Auto-Calibrating Parallel MRI

More Related Content

Similar to An Analytical Comparison of Different Regularization Parameter Selection Methods for Regularization in Auto-Calibrating Parallel MRI (20)

More from SARADINDU SENGUPTA (7)

Recently uploaded (20)

An Analytical Comparison of Different Regularization Parameter Selection Methods for Regularization in Auto-Calibrating Parallel MRI