Fast MRI Reconstruction Using Deep Learning-based

Compressed Sensing: A Systematic Review

Mojtaba Safari1 , Zach Eidex1 Chih-Wei Chang1 , Richard L.J. Qiu1 , Xiaofeng Yang1,‡

1
Department of Radiation Oncology and Winship Cancer Institute, Emory University,
Atlanta, GA 30322, United States of America,
‡
corresponding Author

Author to whom correspondence should be addressed. email: [email protected]

keywords: Compressed sensing (CS), magnetic resonance imaging, MRI reconstruction, fast
MRI, MRI acceleration
running title: DL-based CS-MRI

1
Abstract:

Magnetic resonance imaging (MRI) has revolutionized medical imaging, providing a non-invasive and
highly detailed look into the human body. However, the long acquisition times of MRI present
challenges, causing patient discomfort, motion artifacts, and limiting real-time applications. To
address these challenges, researchers are exploring various techniques to reduce acquisition time
and improve the overall efficiency of MRI. One such technique is compressed sensing (CS), which
reduces data acquisition by leveraging image sparsity in transformed spaces. In recent years, deep
learning (DL) has been integrated with CS-MRI, leading to a new framework that has seen remarkable
growth. DL-based CS-MRI approaches are proving to be highly effective in accelerating MR imaging
without compromising image quality. This review comprehensively examines DL-based CS-MRI
techniques, focusing on their role in increasing MR imaging speed. We provide a detailed analysis of
each category of DL-based CS-MRI including end-to-end, unroll optimization, self-supervised, and
federated learning. Our systematic review highlights significant contributions and underscores the
exciting potential of DL in CS-MRI. Additionally, our systematic review efficiently summarizes key
results and trends in DL-based CS-MRI including quantitative metrics, the dataset used, acceleration
factors, and the progress of and research interest in DL techniques over time. Finally, we discuss
potential future directions and the importance of DL-based CS-MRI in the advancement of medical
imaging. To facilitate further research in this area, we provide a GitHub repository that includes up-
to-date DL-based CS-MRI publications and publicly available datasets -
https://github.com/mosaf/Awesome-DL-based-CS-MRI.

Abbreviations:
bSSFP Balanced steady-state free precession

CS compressed sensing

DC Data consistency

DL Deep learning

DL-based CS-MRI Deep learning-based compressed sensing-magnetic resonance imaging

DTI Diffusion-tensor imaging

DWI Diffusion-weighted imaging

E2E End-to-end
FID Fréchet inception distance

FL Federated learning

FLAIR Fluid-attenuated inversion recovery

GRE Gradient echo

MAE Mean absolute error

MRA Magnetic resonance angiography

2
MRI Magnetic resonance imaging

MSE Mean square error

NMSE Normalized mean square error

PD Proton density

PDFS Proton density fat-suppression

PSNR peak signal-to-noise ratio

R Acceleration rate

RMSE Root mean square error

RSS Root sum of squares

T1c T1-w post-contrast MRI

T1-MPRAGE T1w magnetization-prepared rapid gradient-echo

SSIM Structural similarity index

SWI Susceptibility-weighted imaging

1 Introduction
Magnetic Resonance Imaging (MRI) is a highly effective medical tool that produces high-
quality images of soft tissues in the body. It is widely used in the lesion prognosis and
diagnosis, radiation treatment planning, and follow-up examination. However, the Lancet
Oncology Commission recently highlighted a severe shortage of MRI and other medical
imaging technologies in low-income and middle-income countries (Hricak et al., 2021). This
shortage has resulted in 2.5 million deaths worldwide. The installation of MRI scanners
remains low globally, with only 7 MRI scanners per million people installed as of 2020 (Y.
Liu et al., 2021). This is primarily due to the high cost of installation, operation, and
maintenance. Besides, the daily throughput is limited by the long acquisition time required
for each MRI scan. The long wait time reduces the number of patients that can be seen in a
given day (Murali et al., 2023). Additionally, it increases the likelihood of voluntary and
involuntary patient movements (Safari et al., 2023b), causing motion artifacts that affect the
accuracy of the images produced. The estimated cost of motion artifacts induced by patient
movements is around $364,000 per scanner annually (Slipsager et al., 2020).
MRI requires densely sampled k-space to avoid violating the Nyquist criteria, which results
in longer acquisition times for high-resolution images. To reduce the imaging time, k-space
can be undersampled in the phase encoding direction by increasing the spacing between k-
space lines and, therefore, covering the field of view in a shorter amount of time, as
illustrated in Figure 1 for a Cartesian trajectory. Compressed sensing (CS), also known as

3
compressive sensing or compressive sampling, is a method that aims to reconstruct fully-
sampled k-space from undersampled k-space by exploiting the images’ sparse
representation in a transform domain such as Cosine and Wavelet (Haldar et al., 2010). The
CS algorithms optimize a cost function (1) to iteratively reconstruct the image. However, CS
algorithms are unable to completely reconstruct the high-frequency texture content of
images (Ravishankar and Bresler, 2010), limiting them to acceleration factors between 2.5
and 3 (Guo et al., 2021a). In addition, these iterative techniques inevitably increase
reconstruction time.

Figure 1: The schematic diagrams of the k-space sampling pattern with a Cartesian
trajectory are illustrated for (a) fully sampled data and (b) undersampled 2D MRI data in
phase encoding direction with acceleration rate (R) 2 as well as (c) undersampled 3D MRI
in both slice and phase encoding direction with 𝑅 = 2. The yellow and red circles indicate
sampled and skipped data during the data acquisition.

min = ‖𝚿𝐱‖2
𝑥 , (1)
𝑠. 𝑡. ‖𝐲𝛀 − 𝐄𝛀 𝐱‖22 < 𝜖
where 𝐄𝛀 ∈ ℂ𝑀×𝑁 → ℂ𝑀 given 𝑁 > 𝑀 is the encoding operator. It is composed of a coil
sensitivity map, a Fourier transform, and a sampling map with the specified pattern 𝛺. 𝐲𝛀 ∈
ℂ𝑀 , 𝐱 ∈ ℂ𝑁 , and 𝜖 are the undersampled k-space measurement, fully sampled image, and the
threshold controlling the reconstruction fidelity, which is roughly equal to the expected
noise level, respectively.
Parallel imaging is another approach to reduce image acquisition time. Multiple receiver
coils are placed in different positions within the scanner to independently collect a portion
of the k-space data(Deshmane et al., 2012; Larkman and Nunes, 2007). Each coil is most
sensitive to the area closest to it (shown in Figure 2a) with the sensitivity relationship
encoded in the form of sensitivity maps (shown in Figure 2b). The final image is generated
by combining the individual images obtained from different coils by taking their root sum of
squares (RSS) and weighting them by the corresponding sensitivity maps.

4
Figure 2: (a) Images and (b) sensitivity maps, estimated by the ESPIRiT approach (Uecker
et al., 2014), are illustrated for the first 15 receiver coils. The root sum of squares (RSS)
indicates combined weighted receiver coil images by the corresponding sensitivity maps.

Deep learning (DL) algorithms have garnered significant interest in CS-MRI applications.
These DL models have demonstrated superior reconstruction performance at a higher
acceleration rate compared to traditional non-DL-based CS models (Mardani et al., 2018; Qin
et al., 2018). In light of the rapid advancements in this field, we conduct a systematic review,

5
encapsulating the latest developments in state-of-the-art DL-based CS-MRI models. Table 1
provides a concise summary of the most relevant survey articles, revealing technical
differences compared to our review. We also note that significant progress has been made
in DL-based CS-MRI over the last 2 years, especially with growing interest in diffusion
models and the DC layer, so our review better captures current research directions.

Table 1. Related review papers from the DL-based CS-MRI.

References Year Contributions and technical differences

(Yoon et al., 2023) 2023 This review focuses on accelerated musculoskeletal MRI. It does
not discuss the statistics on the quantitative metrics and
acceleration rates.

(Y. Chen et al., 2022 This review details DL algorithms and provided statistics on
2022) quantitative metrics. However, our comprehensive review, in
addition to those statistics, provided detailed explanations
about MRI imaging, such as k-space trajectories, the
implications of different sampling patterns, and parallel
imaging. Our comprehensive review also discussed the clinical
applications of DL-based CS-MRI and provides insights into
future directions.

(Bustin et al., 2020) 2020 The primary consideration is on DL-based CS-MRI for cardiac
imaging. In addition to being wider in scope, our review
discusses clinical applications of interest to imaging centers,
provides relevant statistics, categorizes the DL method used,
and lists related references.

(Xie and Li, 2022) 2022 A review about CS for medical applications. We focus on CS for MRI
and categorize based on the study’s training method. In addition,
our comprehensive review provides details about MRI acquisition
and acceleration methods.

6
The PubMed database was meticulously searched on February 1st, 2024, using the terms
“deep learning reconstruction”, “fastMRI”, “unrolled optimization”, “MRI reconstruction”,
and “MRI acceleration” for articles published from February 2024 to January 2016. Relevant
studies were carefully screened by title and abstract content. Of the 873 publications
identified by PubMed, 94 articles were included. Figure 3 illustrates the entire literature
screening and selection process.

Figure 3: Flowchart of our study selection process.

1.1 k-space trajectories

Various trajectories have been developed in MRI for traversing k-space, including Cartesian,
spiral, radial, and random trajectories. The Cartesian trajectory, as depicted in Figure 4a, consists
of parallel lines equidistant from each other, with each line representing a frequency-encoding
readout. The image can be reconstructed using a fast Fourier transform, but each line requires a
separate RF pulse, prolonging imaging time.

The radial trajectory, first used by (Lauterbur, 1973) and shown in Figure 4b, consists of spokes
radiating from the center, with an oversampling center in k-space that makes it robust to motion
artifacts (Maclaren et al., 2013). However, undersampling in the azimuthal direction increases
streak artifacts (Xue et al., 2012).

The spiral trajectory shown in Figure 4c was introduced to decrease the MRI acquisition time. It
starts at the center of the k-space and spirals outward, similar to radial sampling, and is robust to
motion artifacts. However, hardware limitations restrict imaging efficiency and increase image
blurring.

7
Figure 4: An exemplary k-space with overlaid (a) Cartesian, (b) radial, and (c) spiral
trajectories illustrated.

1.2 Sampling patterns

To speed up the process of capturing images, a pattern or mask, denoted as 𝛀 in (1), is
utilized to sample k-space. Numerous sampling patterns have been introduced for data
acquisition techniques including Cartesian (Safari et al., 2024), Poisson (Slavkova et al.,
2023), Gaussian (C. Hu et al., 2021), radial (Terpstra et al., 2023), and spiral patterns.
Figure 5 illustrates examples of the different sampling patterns. The Cartesian pattern is
typically used for brain and knee data with a Cartesian k-space trajectory while we found the
Poisson and Gaussian patterns to be primarily used to train self-supervised models as
described in Section 3.4. On the other hand, radial and spiral patterns are mainly used for
capturing myocardial and dynamic images that are more likely to have motion artifacts.

8
Figure 5: Widely used sampling pattern for k-space undersampling - (a) radial, (b) spiral,
(c) Poisson disk, (d) Gaussian, (e) random Cartesian, and (f) equidistant Cartesian.

2 Deep learning
2.1 Convolutional neural network
Convolutional neural networks, also known as ConvNets, are a type of deep neural networks
that are designed to analyze grid-like data such as images and speech (LeCun et al., 1995).
They have gained widespread recognition after the success of AlexNet (Krizhevsky et al.,
2012) and have since been used to achieve state-of-the-art performance in various medical
image processing and analysis tasks. ConvNets typically consist of multiple layers, with each
layer including a convolutional operator, batch normalization layers, nonlinear activation
functions, and dropout layers. Nonlinear activation functions are used to facilitate the
learning of complex functions. Finally, weight regularization and dropout layers are
employed to mitigate overfitting. During convolution, trainable convolution kernels slide
over the images to extract multiple feature maps called channels.
The network’s optimum parameters are computed using a backpropagation algorithm that
calculates the gradient of the cost function with respect to the parameters in each layer.
Batch normalization layers are crucial in training deep ConvNets to prevent vanishing and
exploding gradients. In addition, residual blocks (He et al., 2016) are a popular choice for
building advanced ConvNets due to their ability to prevent gradient vanishing and facilitate
smoother error surfaces (Li et al., 2018). By incorporating skip-layer connections between

9
input and output, residual blocks can help reduce the risk of local minima. Furthermore,
when combined with (batch) normalization layers, residual blocks can effectively address
the problems of vanishing and exploding gradients.

2.2 Network architectures

2.2.1 U-Net
Several deep learning models with different architectures have been proposed to enhance
the performance and generalization of ConvNets. Among them, U-Net, with its elegant design
that utilizes skip connections between the encoder and decoder, is the best known
architecture in computer vision (Ronneberger et al., 2015). It has been extensively exploited
in different medical applications such as image synthesis (Han, 2017), segmentation (Dong
et al., 2019), and registration (Balakrishnan et al., 2019). In recent years, U-net architectures
have incorporated residual and attention layers as a backbone to increase the network’s
depth and improve performance.

2.2.2 Transformer
While ConvNets have been impressive in their results, they are limited by the local context
of convolutional operations. To address this challenge, Transformers have emerged as a
solution to capture global context (Dosovitskiy et al., 2020), often outperforming ConvNets.
However, transformer models are fundamentally very complex and require many trainable
parameters and large databases for training which can be a challenge in medical imaging. To
mitigate these issue, various variations have been proposed, such as Swin Transformers (Z.
Liu et al., 2021), Vision CNN-Transformer (Fang et al., 2022), and ReconFormer (P. Guo et al.,
2023), which aim to reduce model size while improving or maintaining performance.

2.2.3 Generative adversarial network

Generative adversarial networks (GANs) are implicit methods. Thus, they do implicitly
attempt to minimize likelihood function nor attempt to learn latent representation. The GAN,
initially introduced in 2014, consists of two networks, generative and discriminator
(Goodfellow et al., 2020). The former is trained to generate artificial data samples to
approximate the target data distribution, and the latter is simultaneously trained to
distinguish the artificial data from real ground truth data. Thus, the discriminators
encourage the generator to generate data samples with a distribution similar to the target
distribution. Variations of GANs have been developed to perform tasks including image-to-
image translation, such as conditional GAN (Mirza and Osindero, 2014), StyleGAN (Karras et
al., 2019), CycleGAN (Zhu et al., 2017), and Pix2Pix (Isola et al., 2017). GANs are widely used
in medical imaging for tasks such as image registration, image synthesis, and MRI image
reconstruction (Quan et al., 2018; Shaul et al., 2020; Yang et al., 2017).

10
2.2.4 Diffusion model
The stable diffusion model, inspired by nonequilibrium thermodynamics, aims to simplify
complex and difficult-to-calculate distributions using tractable ones like normal Gaussian
distributions (Sohl-Dickstein et al., 2015). This model is comprised of two steps - the forward
and reverse processes (Figure 6). During the forward process, Gaussian noise is added to the
initial image 𝑥0 over 𝑇 steps until the data at step 𝑇 becomes normal Gaussian noise 𝑥𝑇 =
𝒩(𝟎, 𝐈). In the reverse process, the model learns to recover the original image 𝑥0 from its
noisy version given at a step 𝑡 ∈ (0, 𝑇] (Chan, 2024).

Figure 6: Forward and reverse diffusion processes. The first row indicates data in image
space, and the second row indicates the corresponding data distribution. The forward
diffusion process adds Gaussian noise in 𝑇 steps in a controlled way to produce normal
Gaussian noise in step 𝑇. The reverse diffusion process requires a model parametrized by 𝜑
to learn input 𝑥0 from noise-corrupted image 𝑥𝑡 in a given step 𝑡.

Stable diffusion models have been employed for medical imaging tasks such as denoising
(Pan et al., 2023b), synthesis (Pan et al., 2023a), MRI distortion reduction (Safari et al.,
2023b), and MRI image reconstruction (Chung and Ye, 2022; Güngör et al., 2023).

11
3 Deep learning for MRI reconstruction
The framework for DL-based CS-MRI can be divided into two main categories: data-driven
and physics-driven models. Within these categories, there are two types of models: end-to-
end and unroll. End-to-end models take in zero-filled k-space and output fully-sampled k-
space. They typically utilize a regularization term listed in Table 2 to enforce the
uniqueness of the reconstructed images. On the other hand, unroll models are more
complex and further classified into two types: unroll optimization and closed-form models
known for the “data consistency (DC) layer.” Unroll optimization models iteratively
optimize the reconstruction process, while DC layer models use a closed-form equation to
ensure data consistency. These models are utilized in various training scenarios, including
federated learning and self-supervised training. For a comprehensive understanding of DL-
based CS-MRI methods and their corresponding components and features, please refer to
Figure 7.

Figure 7: An overview of seven categories of DL-based CS-MRI methods.

Figure 8 shows a stacked chart of the number of publications since 2018 by category. The
total number of publications has grown exponentially in recent years and interest in the DC
layer method continues to increase while interest in unrolling optimization and end-to-end
approaches remains consistent over time.

12
Figure 8: Overview of publications in DL-based CS-MRI over time. The dashed line indicates
the general trend plotted using 1 + exp(0.62 × 𝑡) where t is defined in years.

3.1 End-to-end models

DL end-to-end models are specifically designed to tackle the CS-MRI problem without
enforcing any data acquisition model. To achieve this, these models rely on a neural network
to accurately predict fully sampled data from undersampled data (B. Levac et al., 2023;
Mardani et al., 2018; Yang et al., 2016). Additionally, these models are trained using various
regularization techniques that help address the ill-posed inverse problem. Table 2 lists
common regularization techniques utilized in these models.

Table 2: The widely used regularization terms in deep CS-MRI are summarized.

Traditional DL-based model Regularizer Reference

Dictionary learning ADMS ‖𝛼‖1 (Cao et al., 2020)

Field of expert VN ∑ 〈Φ𝑖 (𝐾𝑖 𝑥),1〉 (Hammernik et al.,

𝑖 2018)

pISTA-SENSE pISTA-SENSE- ‖Ψ𝑥‖1 (Lu et al., 2020)

ResNet

Total variation -, RELAX ‖∇𝑥‖1 (F. Liu et al., 2021;

Sun et al., 2017)

13
 Sparse and low-rank ODLS ‖ℋ𝑥‖∗ (Z. Wang et al.,
model 2022)

Dictionary learning learns a latent representation 𝛼 of the input image 𝑥 where the ℓ1
norm enforces it to be sparse.
Field of expert is composed of the convolution kernel 𝐾𝑖 and Φ𝑖 , which they are learned
from data.
pISTA-SENSE is a projected iterative soft-thresholding algorithm that solves (1)
iteratively where the transform Ψ enforces the sparsity of the reconstructed image 𝑥.
Total variation enforces image smoothness by minimizing the image gradient
variations.
Sparse and low-rank model minimize the nuclear norm ‖ℋ𝑥‖∗ where ℋ is the Hankel
matrix.

The end-to-end approach employs the same baseline models that are used for image-to-
image translation, such as the U-nets (Hyun et al., 2018; Xiang et al., 2018), Swin
transformers (Huang et al., 2022), and GANs (Shitrit and Riklin Raviv, 2017; Zhao et al.,
2023). However, they require a larger sample dataset than unrolling CS-MRI models and
tend to predict images with synthetic data.. Table 3 provides a list of selected references that
used end-to-end DL models to solve DL-based CS-MRI algorithms.

Table 3: Overview of supervised end-to-end models to predict the fully sampled images.

sample
DL Model Dataset Region Modality R Ref. Code1
size
fastMRI2
(Narnhofer et
GAN (Zbontar et Knee PDFS,PD 299,300 4,6 No
al., 2019)
al., 2018)
IXI3 T2,PD 72
(Lyu et al.,
GAN Brain 2,3,4 No
2020)
NAMIC4 T1,T2 20
(Hammernik (Lu et al.,
ResNet knee PD 15 5,7,9 Yes
et al., 2018) 2020)
T1𝜌-map, T2- (Li et al.,
U-net Institutional5 Knee 10 2,4,5,6 No
map 2023)
MICCAI
Brain T1 150 (Gao et al.,
GAN 20136 2,3.3,10 No
2023)

14
 ASC7 Cardiac T1c 100

fastMRI Knee PD 96
(Do et al.,
Y-net Own Brain T1,T2 37 2,3,4,5,6,8 No
2020)
(Ayde et al.,
U-net Own Knee T1 10 3,4,5 No
2022)
IXI
Brain T1 573
MICCAI (Zhao et al.,
SwinGAN Brain NS8 24809 20,30,50 Yes
2013 2023)
Knee NS 24809
MRNet
Dense- MICCAI (Xiang et al.,
Brain T2 5 2,4,8 No
Unet 2016 2018)
(J. Wang et al.,
U-net Institutional Cardiac bSSFP,GRE 80 NS No
2024)
HCP T1,T2 20
GAN (Meng et al.,
Brain 8 No
(Pix2pix) 2021)
Institutional T2-map NS
1
Cardiac (Liu et al.,
GAN Institutional 0 DTI 30 2.6 No
2023)
DeepADC Animal (Li et al.,
Institutional DWI 183 4,8 Yes
-Net study 2024)
Brain, 5847,11 (Khawaled and
NPB-REC fastMRI NS 4,8,12 Yes
Knee 67 Freiman, 2024)
1.78,2.39,
2.91,
T1,T2,T2FLAI 3.33,3.64, (Jin and Xiang,
SCU-Net Institutional Brain 180 No
R 4, 2023)
4.27,4.44,
4.71
DuDReTL (Hong et al.,
Uk Biobank11 Cardiac NS 9032 20 Yes
U-net 2023)

fastMRI T1,T2,T1c,T2F 10
(Dawood et al.,
iRAKI Brain LAIR 4,5 Yes
2023)
Institutional 3
T1,T2,T1FS
fastMRI Brain T2FLAIR 10009
MA- (Avidan and
4,8 Yes
RECON Freiman, 2024)
MRIdata12 Knee PD 418729

15
Diffusion (Cao et al.,
fastMRI Knee PD 973 10,12 Yes
model 2024)
1
Whether the original implementation is publicly available,
2
https://fastmri.med.nyu.edu/,
3
https://brain-development.org/ixi-dataset/,
4
https://insight-journal.org/,
5
The study used its Institutional dataset,
6
MICCAI 2013 grand challenge dataset:
https://wiki.cancerimagingarchive.net/display/Public/NCI-
MICCAI+2013+Grand+Challenges+in+Image+Segmentation,
7
Atrial Segmentation Challenge: https://www.cardiacatlas.org/atriaseg2018-challenge/,
8
Information not available from the publication,
9
Number of image slices was specified,
10
Ex-vivo study
11
https://www.ukbiobank.ac.uk/
12
https://www.mridata.org/

3.2 Unroll model

3.2.1 Unroll optimization
Unroll CS-MRI models combine DL with a data acquisition model to solve an optimization
problem iteratively. This optimization is given by Equation 2:

⏟ Ω − 𝐸Ω 𝑥‖22 + 𝜆
𝑥̂ = 𝑎𝑟𝑔𝑚𝑖𝑛 ‖𝑦 ℛ(𝑥)
⏟ (2)
𝐷𝑎𝑡𝑎 𝑐𝑜𝑛𝑠𝑖𝑠𝑡𝑒𝑛𝑐𝑦 𝑅𝑒𝑔𝑢𝑙𝑎𝑟𝑖𝑧𝑎𝑡𝑖𝑜𝑛

In this equation, λ > 0 is a scalar regularization weight that balances between the data
consistency and regularization terms. The data consistency term ensures that the
reconstructed images x̂ are similar to the given undersampled yΩ , thus enforcing data
fidelity. The regularization term helps solve the ill-posed CS-MRI problem by imposing
sparsity on the solution to guarantee the uniqueness of the reconstructed images, 𝐱̂ (Donoho,
2006).

16
DL models, particularly ConvNets, are used heavily to learn the regularization term through
an unroll training scheme. Similarly, the prior DL-based regularizes encode prior knowledge
about the reconstructed images, such as sparsity. The unroll CS-MRI DL models have shown
to outperform end-to-end methods using a network with a smaller number of trainable
parameters (Aggarwal et al., 2018; Geng et al., 2023; Liu et al., 2022; Qiao et al., 2023; Qu et
al., 2024).
However, the iterative nature of the unroll optimization method may increase the
computation time during both training and inference steps as both the network’s weights
and data consistency terms are simultaneously updated. Table 4 provides a list of references
that used unroll optimization models to solve DL-based CS-MRI algorithms.

Table 4: Overview of supervised unroll optimization to predict the fully sampled images.

Referen
DL Model Dataset Region modality sample size R Code
ce
IXI Brain T1,T2 50 (R. Liu
2.5, 3.3,
ConvNet et al., Yes
5
fastMRI Knee PD 1025 2020)
Brain T1 5 (D. Guo
Institutional
ConvNet 4,5,6.7 et al., No
Knee PD, PDFS 60 2023)
fastMRI
(Huang
RG-Net Institutional Brain T1𝜌 8 17 et al., No
2017)
(Duan
ConvNet Institutional Brain SWI 117 5,8 et al., No
2022)
HCP (Van (Zufiria
ConvNet Essen et al., Brain T1w,T2w 1200 NS1 et al., No
2013) 2022)
(Zhou
Cardiac,
ConvNet Institutional NS 20,16 3,4,5 et al., No
Abdominal
2019)
fastMRI
T1c, T1,
Brain T2,T2- 5846
fastMRI (Yiase
FLAIR
RecurrentV mis et
Knee 1172 2,4,6 No
arNet Calgary- al.,
PD
Campinas 2024)
Brain 67
(Beauferris
T1w
et al., 2022)

17
 (Hammerni
(Hamm
k et al., Knee PD 20
ernik et
ConvNet 2018) 3,4,6 No
al.,
Brain T1 8
2018)
Institutional
(K.
fastMRI Knee PDFS, PD 380 5
Wang
UFLoss Yes
et al.,
MRIdata Knee PDFS, PD 20 8
2022)
SRI24
(Rohlfing et
al., 2010)
PD,T1,T2
36
(Sun et
MRBrainS1 T1,T2-
DFSN Brain 20 5, 10 al., No
3(Mendrik FLAIR
2019)
et al., 2015)
175
T1,T2
NeoBrainS1
2 (Išgum et
al., 2015)
(Slavko
8,12,18, va et
ConvNet Institutional Brain T1map 3 Yes
36 al.,
2023)
(Ryu et
T1,T2,PD,T2
JDL Institutional Brain 8 4,8 al., Yes
FLAIR
2021)
(Qu et
pFISTA-DR Institutional Brain T1,PD 200 5,7,10 al., No
2024)
(Qiao et
U-net fastMRI Knee PD, PDFS 20 4,6 al., Yes
2023)
(Prama
fastMRI 120
T1,T2,T2- 1.8,2.5, nik et
ConvNet Brain No
FLAIR 3.5,4 al.,
Institutional 3
2023)
T1,T2
NAMIC2 Brain 5 (X. Liu
ConvNet 8 et al., No
T2,T2-
MRBrainS3 Brain 7 2021)
FLAIR

18
 (Kofler
Dictionary
Institutional Cardiac - 19 2,4,8 et al., Yes
learning
2023)
(Kim
PD-PCG- and
fastMRI Knee PD, PDFS 484,489 4 No
Net Chung,
2022)
(Y. Hu
4
RUN-UP Institutional Brain, Breast DTI, DWI 14,6 100 et al., No
2021)
(Sandin
10,12,1
ConvNet Institutional Cardiac NS 22 o et al., No
4
2021)
(Xu et
CEST-VN Institutional Brain CEST 54 3,4,5 al., Yes
2024)
(Safari
Diffusion 1.25,1.6
GLIS-RT Brain T1c 230 et al., No
model 6,2.5,5
2024)

1
It is not specified in the manuscript
2
http://hdl.handle.net/1926/1687
3
https://mrbrains18.isi.uu.nl/

3.2.2. Data consistency layer

The more popular approach is to train the unrolled models similarly to the end-to-end
models as follows:

1
⏟ Ω − 𝐸Ω 𝑥‖22 ,
𝑥̂𝑓𝜓 = 𝑎𝑟𝑔𝑚𝑖𝑛 ‖𝑥 − 𝑓𝜓 (𝑥Ω )‖ + 𝜆 ‖𝑦 (3)
𝐷𝑎𝑡𝑎 𝑐𝑜𝑛𝑠𝑖𝑠𝑡𝑒𝑛𝑐𝑦

where 𝑓𝜓 is a DL model that maps the undersampled input images 𝑥𝛺 to reconstruct fully
sample images 𝑥. The DL reconstruction and data consistency operate on the image domain
and k-space domain, respectively. Although the DL part is trained without incorporating a
priori information, the second term discourages the DL first part from updating the k-space
parts that were not sampled (Schlemper et al., 2017). The closed form for (3) is as follows
(Qin et al., 2018):

19
 𝑥̂𝑓𝜓 (𝑘) 𝑖𝑓 𝑘 ∉ Ω
𝑥̂ = {𝑥̂𝑓 (𝑘) + 𝜆0 𝑥Ω (𝑘) (4)
𝜓
𝑂𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒
𝜆0 + 1
This closed form is a computational layer called the DC layer at the end of DL models. The DC
layer is a crucial part of the CS-MRI DL model, playing an important role in reconstructing
images (Cheng et al., 2021; Korkmaz et al., 2023). The DC layer allows for a flexible design of
the DL model when it is added to U-net (Murugesan et al., 2021), transformers (Wu et al.,
2023), stable diffusion model (Cao et al., 2024), and so on. Table 5 summarizes the DL-based
CS-MRI trained under the DC layer framework.

Table 5: Overview of supervised models that used the data consistency to predict the fully
sampled images.

sample Cod
DL Model Dataset Region Modality R Ref.
size e
Institution 2,2.5,3.3,5,
(Wu et
KTMR Brain T1c, MRA 17 Yes
al 10al., 2023)
fastMRI Brain T1 973 (B. Wang
DCT-net 4,8
et al., No
Calgary Knee PD 25 2024)
Institution (Qin et
CTFNet Cardiac bSSFP 48 8,16,24 Yes
al al., 2021)
Institution 3.3,4,5,6.7, (Q. Liu et
EDAEPRec Brain T2 7 No
al 10 al., 2020)
Institution T1MPRAGE,T2FLAIR 30,50, (Dai et
GFN Brain 3.3,5,10 No
al ,TOF 80 al., 2023)
Knee PD, PDFS 973 (E. Z.
PC-RNN fastMRI 4,6 Chen et Yes
Brain T1,T2 4469 al., 2022)
Institution (Wu et
ResNet Unet Knee T1 360 6 No
al al., 2019)
Knee, (Wen et
CNF fastMRI PD, T2 20,8 4 Yes
Brain al., 2023)
Institution Abdom (Wang et
stDLNN GRE 8 4,8,16,25 Yes
al en al., 2023)
(Terpstra
Institution 3.7,7.4,14.
MODEST Cardiac 28 et al., Yes
al 8
2023)
Institution Brain T1,T2,PD 22
Deepcomplex (Wang et
al 4,5,10 No
MRI al., 2020)
(Hammern Knee PD 20

20
 ik et al.,
2018)
fastMRI T1, T2, PD 66 (Güngör
AdaDiff Brain 4,8 et al., Yes
IXI T1,T2,T2-FLAIR 420 2023)
(Ekanaya
584,
McSTRA fastMRI Knee PD, PDFS 4,6,8,10,12 ke et al., No
588
2024)
Cardiac Cardiac
Cardiac, 200
,
Kirby
T1-MPRAGE 42
Brain,
(Muruges
Calgary
DC-RSN T1 45 4, 5 an et al., No
Brain,
2021)
MRBrain
T2-FLAIR 7
Brain,
Hammerni
PDF, PD, T1, T2, T2FS 25
ck Knee

3.3 Federated learning

Federated learning (FL) is a promising framework that enables the collaborative training of
learning-based models across multiple institutions without the need for sharing local private
data (Yang et al., 2019). The objective of FL models is to learn a global model by taking the
average of local models (McMahan et al., 2017) or by ensuring the proximity of local models
to the global model (Li et al., 2020). When applied to MRI image reconstruction, the FL offers
unique advantages tailored to the specific challenges and requirements as follows:
• MR images often contain sensitive patient information that needs to be protected. FL
enables MRI models to be trained directly on the devices where the images are acquired,
without the need to transmit patient data to a centralized location. This decentralization
of data ensures privacy and confidentiality of patient information is maintained.
• MRI machines can vary in their hardware specifications and imaging protocols, which
can lead to challenges in standardizing image reconstruction algorithms. However, FL
accommodates this heterogeneity by allowing models to be trained collaboratively
across different types of MRI machines, ensuring that the reconstruction algorithms are
robust and adaptable to various configurations.

It is worth noting that FL models are predominantly supervised and have been developed
under the end-to-end and unroll model frameworks, which have shown promising results in
various applications.

21
Table 6. Overview of supervised models that were trained with federated learning.

DL Training
Dataset sample size Ref. Code
Model framework
fastMRI,
3443
HPKS(Jiang et al., 2019)
144
FL-
End-to-End (Guo et al., 2021b) Yes
MRCM IXI,
434
BraTS(Menze et al.,
494
2014)
fastMRI 2134

BraTS 385
FedMRI End-to-End (Feng et al., 2022) Yes
SMS (Feng et al., 2021) 155

uMR (Feng et al., 2021) 50

fastMRI 3443

BraTS 494
ACM-
DC guided (Lyu et al., 2023) No
FedMRI
IXI 526

Institutional 150
fastMRI 51

BraTS 55
FedGIM
DC guided (Elmas et al., 2022) Yes
P
IXI 55

Institutional 10
Unrolled (B. R. Levac et al.,
Unroll fastMRI, MRIdata NS Yes
FedLrn 2023)

3.4 Self-supervised learning

In contrast to supervised learning methods that necessitate fully sampled ground truth
images, self-supervised models alleviate this requirement and are often trained using
unrolling optimization techniques. The training framework of self-supervised algorithms
does not mandate fully sampled ground truth images. This approach is particularly
advantageous in scenarios where obtaining fully sampled data without distortions is

22
challenging, such as myocardial perfusion with the patient’s involuntary movements, which
cause motion artifacts (Haji-Valizadeh et al., 2018). Self-supervised methods draw samples
from the undersampled pattern 𝛺 provided in (3) to generate a new pattern 𝛬 and 𝛩, where
𝛺 = 𝛬 ∪ 𝛩 and 𝛩 = 𝛺 ∖ 𝛬. The former is used to train the DL model, while the latter is
utilized to compute the loss (Feng et al., 2023; Heydari et al., 2024).

Table 7: Overview of self-supervised models to predict the fully sampled images.

Regio sample
DL Model Dataset modality R Ref. Code
n size
IXI Brain T1, T2, PD 473
DC-
4, 5 (Yan et al., 2023) No
SiamNet
MRINet Knee PD 1250

PDFS
MRIdata Knee 19
Noise2Reco (Desai et al.,
12,16 Yes
n T2,T2- 2023)
fastMRI Brain 603
FLAIR,T1,T1c

(Zhou et al.,
DDSS HCP Brain T1 505 2,4 No
2022)
Institutio Brain T1MPRAGE 19
nalfastM
(Yaman et al.,
SSDU RI Knee PD,PDFS 35,25 4,6,8 Yes
2020)
fastMRI Brain T2 10
Institutio
Brain T1MPRAGE 9
Multi-mask nal (Yaman et al.,
8,12 No
SSDU MRIDdat 2022)
Knee PD 20
a
Institutio Brain, (F. Liu et al.,
RELAX T1 and T2 maps 21 5 No
nal Knee 2021)
Joint Institutio (Heydari et al.,
Brain T1 and T2* maps 2 16,25 Yes
MAPLE nal 2024)

3.5 Assessment
In the previous sections, we covered three major DL-based CS-MRI training approaches: end-
to-end, unroll optimization, and the DC layer. While these methods can be applied to both
supervised and self-supervised frameworks, the unroll optimization and DC layer
approaches are typically used for self-supervised training. For your convenience, we
summarized the advantages and disadvantages of each approach in Table 8.

23
Table 8: A bullet list of the pros and cons of each training framework is provided.

DL method Pros Cons

1. Requires a big dataset
2. Likely to add unwanted image
1. Free from MRI data acquisition
structure, specifically for higher
method
End-to-end acceleration rate
2. Easy to employ models proposed for
3. Limited generalization to out-of-
image synthesis and segmentation
distribution data

1. Iterative method
1. Requires smaller datasets compared
2. Requires image acquisition
with the end-to-end method
Unroll knowledge
2. More likely to generalize better
optimization 3. Requires undersampling pattern in
3. Performs well using a ConvNet with
inference time
small number of trainable parameters
1. Requires smaller dataset than the
1. Requires undersampling pattern in
end-to-end method
DC Layer inference time
2. Provides closed form equation
3. Easy to implement

4 Evaluation metrics
4.1 Quantitative metrics
When testing CS-MRI models, the quality of reconstructed images is quantitatively assessed
by comparing them with the ground truth. Retrospective undersampling of the k-space
allows the reporting of indices to quantify the quality of the reconstructed images.

24
Figure 9: This graph illustrates (a) The metric used in different studies, (b) the training
method used in different studies, (c) the dataset used for training used in the studies, and (d)
the acceleration rate of the undersampling. Abbreviations: SSIM: structural similarity index;
PSNR: peak signal-to-noise ratio; NMSE: normalized mean square error; NRMSE: normalized
root mean square error; MSE: mean square error; MAE: mean absolute error; RMSE: root
mean square error; and FID: Fréchet inception distance.

Most studies quantitatively compare predicted images with ground truth images. As
indicated in Figure 9a, most of these studies use the structural similarity (SSIM) index or
peak signal-to-noise ratio (PSNR) to compare reconstructed image 𝑥̂ with ground truth
image 𝑥 as follows:
L2
PSNR(x̂, x) = log10 , (5)
1
∥ x̂ − x ∥22
N
where ∥⋅∥2 is the squared Euclidean distance, 𝑁 is the number of images’ voxels, and 𝐿 is the
maximum voxel intensity of 𝑥. A higher PSNR indicates a better reconstruction. The
logarithmic operator quantifies image quality that closely aligns with human perception
(Safari et al., 2023a).

(2μx̂ μx + c1 )(2σx̂x + c2 )
SSIM(x̂, x) = , (6)
(μ2x̂
+ μ2x + c1 )(σ2x̂ + σ2x + c2 )

25
where 𝜇𝑥̂ and 𝜇𝑥 are the average voxel intensities in 𝑥̂ and 𝑥, 𝜎𝑥̂ and 𝜎𝑥 are the variance, and
𝜎𝑥̂𝑥 is the covariance between 𝑥̂ and 𝑥. The constants 𝑐1 and 𝑐2 stabilizes the division, which
usually are 𝑐1 = (𝑘1 𝐿)2 and 𝑐2 = (𝑘2 𝐿)2 . SSIM ranges between -1 and 1, with the best
similarity achieved by an SSIM equal to one.
The normalized mean square error (NMSE) has become more popular since 2022 to quantify
the quality of reconstructed images. The NMSE is defined as
∥ x̂ − x ∥22
NMSE(x̂, x) = , (7)
∥ x ∥22
Smaller NMSE values indicate better image reconstruction. However, it favors image
smoothness rather than sharpness.
However, other metrics such as root mean square error, mean square error, mean absolute
error, and Fréchet inception distance are rarely used, especially after a recommendation
made in 2018 by Zbontar, Jure, et al. (Zbontar et al., 2018) (see Figure 8a).
Regarding the training methods, the Unroll models, including the Unroll optimization and DC
layer, are the most commonly used, with a growing use of the DC layer since 2020, which is
expected to continue this trend in the future. More details about these trends are illustrated
in Figure 9b.
Our systematic review found that around 46% of studies used their own dataset, while
fastMRI was the most frequently used public dataset. The majority of private datasets use
single-coil images compared to the later’s raw multi-coil raw 2D k-space data. The fastMRI
dataset consists of three imaging regions: the brain, pelvis, and knee regions. Less commonly
used datasets include IXI, MRIdata, Calgary, and MICCAI challenges, with only around 7%,
5%, and 3.5% usage rates, respectively (see Figure 9c).
In addition, most studies reviewed by this study tested their proposed model using
acceleration factor (R) 2 ≤ 𝑅 < 6. The least simulated acceleration factor was 𝑅 ≥ 12 and
6 ≤ 𝑅 < 8 with 13.4% and 7.3% of usage, respectively. The percentage of acceleration usage
is summarized in Figure 9d.

4.2 Clinical evaluations

The metrics presented in Section 4.1 quantify the quality of image reconstruction, but their
results may not directly correlate with clinical outcomes. Several studies have been
conducted to evaluate the clinical significance of CS-MRI using DL models. For example, a
study found that DL-based CS-MRI and fully sampled MRI images showed no significant
differences (p-values > 0.05) in the organ-based image quality of the liver, pancreas, spleen,
and kidneys, number and measured diameter of the detected lesions while reducing the
imaging time by more than 85% (Herrmann et al., 2023).
Similarly, another study showed that brain MRI images accelerated up to 4 × and 14 × had
sufficient image quality for diagnostic and screening purposes, respectively (Radmanesh et

26
al., 2022). A third study found that there was no statistical significance (p-value = 0.521)
between the DL-based T2-FLAIR MRI image and standard T1c MRI images in the assessment
of inflammatory knee synovitis (Feuerriegel et al., 2023). These studies are consistent with
another conducted to compare the diagnostic performance of DL MRI and standard MRI
images in detecting knee abnormalities (Johnson et al., 2023).
A recent study found that there is no significant difference in the overall quality of MRI
images generated by DL and standard fully sampled images for various MRI sequences,
including T2 and diffusion-weighted imaging, for patients with prostate cancer. The study
also revealed that DL MRI and standard MR images identified a similar number of Prostate
Imaging Reporting and Data System (PIRADS) ≥ 3 lesions. However, the imaging time was
significantly reduced by about 3.7-fold with the use of DL MRI. This study’s findings suggest
that DL MRI can be a viable alternative to standard MRI imaging for prostate cancer patients,
as it can produce similar quality images in a significantly shorter acquisition time (Johnson
et al., 2022).

5 Discussion
The rapid advancement of DL in the field of computer vision has led to a significant increase
in the number of studies utilizing DL to solve CS-MRI, as depicted in Figure 7. In this study,
we provide a comprehensive overview of DL models and training approaches, including the
use of GAN and neural architecture such as U-net and vision transformers for an end-to-end
approach. Recent studies have extensively used the unroll optimization and DC layers due to
their advantages, allowing for the employment of smaller ConvNets. This, in turn, reduces
the number of trainable parameters, minimizing the risk of overfitting. Moreover, our study
also revealed that the majority of studies employ SSIM, PSNR, and NMSE metrics to measure
the image quality of predicted images.
The objective of DL-based CS-MRI approaches is to decrease imaging time, thereby
enhancing throughput and minimizing the chances of patient movement. Apart from this,
other methods, such as super-resolution and image synthesis, can also be adopted to
decrease imaging time. These approaches strive to enhance the resolution of images
obtained from lower spatial resolution (Chang et al., 2024; Xie et al., 2022) and produce MRI
images of high spatial resolution and signal-to-noise ratio from lower B0 (Eidex et al., 2023).
These techniques can significantly improve MRI imaging with permanent magnets having
low B0 , ultimately increasing the image quality. Portable MRI scanners widely use
permanent magnets, which substantially reduce maintenance costs and make them more
suitable for low-income and middle-income countries. Although these techniques can reduce
the need for stronger gradient magnets and minimize patients’ nerve stimulation, they were
not included in this systematic review since they did not employ compressed sensing
algorithms to train their models.
The coil sensitivity map, depicted in Figure 2b, is essential to consider the non-uniform
sensitivity of the receiver coils. A precise sensitivity map is crucial for generating consistent
and accurate MRI images and quantitative MRI maps across different hospitals. Most of the
reviewed studies use the ESPIRiT algorithm (Uecker et al., 2014) provided by the Berkeley

27
Advanced Reconstruction Toolbox (https://mrirecon.github.io/bart/) to pre-calculate the
sensitivity map. However, this algorithm involves significant computation that may hinder
its application in MRI-guided surgery and treatment, where rapid image reconstruction is
required. We anticipate that DL-based CS-MRI techniques, which can simultaneously predict
coil sensitivity maps and coil images (Feng et al., 2023; Sun et al., 2023; Z. Wang et al., 2024),
will be explored further in the future, particularly for MRI-guided treatment methods such
as MRI-guided adaptive radiation therapy. Furthermore, most CS-MRI methods require prior
knowledge of the sampling pattern, which may not be available to the user in real clinical
applications. Therefore, an approach that can handle deviations from the actual sampling
mask or jointly predict the optimal sampling pattern and predict the images (Seo et al., 2022)
is required.
In a recent systematic review by Hu, Mingzhe, et al., the use of language models in medical
imaging was explored in detail (Hu et al., 2024). However, our comprehensive review has
found that the utilization of foundational models in training DL-based CS-MRI is currently
quite limited. We believe that these models could have a more significant impact on DL-based
CS-MRI. By providing prior knowledge about the sampling pattern without explicitly
specifying the sampling images, these models could be extremely helpful. Additionally, their
prior inputs about the MRI sequence and imaging region could aid in training a model
specific to that region and sequence, potentially improving the reconstruction of out-of-
distribution and out-of-region images.

6 Concluding remarks
The current DL-based CS-MRI models are usually trained by using 2D fastMRI datasets,
limiting the spatial resolution for small lesion detection. With the advancement of MRI
techniques, future research can significantly benefit from the availability of large 3D or 4D
raw k-space datasets featuring abnormalities. Such datasets can enable the development of
tailored 3D models, potentially allowing real-time tumor tracking during radiation therapy
for patients with conditions like lung cancer. By incorporating high-dimensional datasets, it
would be possible to report accurate clinical endpoints to enhance cancer prognosis.

Acknowledgments
This research is supported in part by the National Institutes of Health under Award Number
R56EB033332, R01EB032680, and R01CA272991.

Conflicts of interest
There are no conflicts of interest declared by the authors.

28
References

Aggarwal, H.K., Mani, M.P., Jacob, M., 2018. MoDL: Model-based deep learning architecture
for inverse problems. IEEE transactions on medical imaging 38, 394–405.
Avidan, N., Freiman, M., 2024. MA-RECON: Mask-aware deep-neural-network for robust
fast MRI k-space interpolation. Computer Methods and Programs in Biomedicine
244, 107942.
Ayde, R., Senft, T., Salameh, N., Sarracanie, M., 2022. Deep learning for fast low-field MRI
acquisitions. Scientific reports 12, 11394.
Balakrishnan, G., Zhao, A., Sabuncu, M.R., Guttag, J., Dalca, A.V., 2019. Voxelmorph: a
learning framework for deformable medical image registration. IEEE transactions
on medical imaging 38, 1788–1800.
Beauferris, Y., Teuwen, J., Karkalousos, D., Moriakov, N., Caan, M., Yiasemis, G., Rodrigues, L.,
Lopes, A., Pedrini, H., Rittner, L., others, 2022. Multi-coil mri reconstruction
challenge—assessing brain mri reconstruction models and their generalizability to
varying coil configurations. Frontiers in Neuroscience 16, 919186.
Bustin, A., Fuin, N., Botnar, R.M., Prieto, C., 2020. From Compressed-Sensing to Artificial
Intelligence-Based Cardiac MRI Reconstruction. Front. Cardiovasc. Med. 7, 17.
https://doi.org/10.3389/fcvm.2020.00017
Cao, C., Cui, Z.-X., Wang, Y., Liu, S., Chen, T., Zheng, H., Liang, D., Zhu, Y., 2024. High-
Frequency Space Diffusion Model for Accelerated MRI. IEEE Transactions on
Medical Imaging.
Cao, J., Liu, S., Liu, H., Lu, H., 2020. CS-MRI reconstruction based on analysis dictionary
learning and manifold structure regularization. Neural Networks 123, 217–233.
Chan, S.H., 2024. Tutorial on Diffusion Models for Imaging and Vision. arXiv preprint
arXiv:2403.18103.
Chang, C.-W., Peng, J., Safari, M., Salari, E., Pan, S., Roper, J., Qiu, R.L., Gao, Y., Shu, H.-K., Mao,
H., others, 2024. High-resolution MRI synthesis using a data-driven framework with
denoising diffusion probabilistic modeling. Physics in Medicine & Biology 69,
045001.
Chen, E.Z., Wang, P., Chen, X., Chen, T., Sun, S., 2022. Pyramid convolutional RNN for MRI
image reconstruction. IEEE Transactions on Medical Imaging 41, 2033–2047.
Chen, Y., Schonlieb, C.-B., Lio, P., Leiner, T., Dragotti, P.L., Wang, G., Rueckert, D., Firmin, D.,
Yang, G., 2022. AI-Based Reconstruction for Fast MRI—A Systematic Review and
Meta-Analysis. Proc. IEEE 110, 224–245.
https://doi.org/10.1109/JPROC.2022.3141367
Cheng, J., Cui, Z.-X., Huang, W., Ke, Z., Ying, L., Wang, H., Zhu, Y., Liang, D., 2021. Learning
data consistency and its application to dynamic MR imaging. IEEE Transactions on
Medical Imaging 40, 3140–3153.
Chung, H., Ye, J.C., 2022. Score-based diffusion models for accelerated MRI. Medical image
analysis 80, 102479.
Dai, Y., Wang, C., Wang, H., 2023. Deep compressed sensing MRI via a gradient-enhanced
fusion model. Medical Physics 50, 1390–1405.
Dawood, P., Breuer, F., Stebani, J., Burd, P., Homolya, I., Oberberger, J., Jakob, P.M., Blaimer,
M., 2023. Iterative training of robust k-space interpolation networks for improved

29
 image reconstruction with limited scan specific training samples. Magnetic
Resonance in Medicine 89, 812–827.
Desai, A.D., Ozturkler, B.M., Sandino, C.M., Boutin, R., Willis, M., Vasanawala, S., Hargreaves,
B.A., Ré, C., Pauly, J.M., Chaudhari, A.S., 2023. Noise2Recon: Enabling SNR-robust
MRI reconstruction with semi-supervised and self-supervised learning. Magnetic
Resonance in Medicine 90, 2052–2070.
Deshmane, A., Gulani, V., Griswold, M.A., Seiberlich, N., 2012. Parallel MR imaging. Magnetic
Resonance Imaging 36, 55–72. https://doi.org/10.1002/jmri.23639
Do, W.-J., Seo, S., Han, Y., Ye, J.C., Choi, S.H., Park, S.-H., 2020. Reconstruction of
multicontrast MR images through deep learning. Medical physics 47, 983–997.
Dong, X., Lei, Y., Wang, T., Thomas, M., Tang, L., Curran, W.J., Liu, T., Yang, X., 2019.
Automatic multiorgan segmentation in thorax CT images using U-net-GAN. Medical
physics 46, 2157–2168.
Donoho, D.L., 2006. For most large underdetermined systems of linear equations, the
minimal ℓ1 solution is also the sparsest solution. Communications on pure and
applied mathematics 59, 797–829.
Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T.,
Dehghani, M., Minderer, M., Heigold, G., Gelly, S., others, 2020. An image is worth
16x16 words: Transformers for image recognition at scale. arXiv preprint
arXiv:2010.11929.
Duan, C., Xiong, Y., Cheng, K., Xiao, S., Lyu, J., Wang, C., Bian, X., Zhang, J., Zhang, D., Chen, L.,
others, 2022. Accelerating susceptibility-weighted imaging with deep learning by
complex-valued convolutional neural network (ComplexNet): validation in clinical
brain imaging. European Radiology 32, 5679–5687.
Eidex, Z., Wang, J., Safari, M., Elder, E., Wynne, J., Wang, T., Shu, H.-K., Mao, H., Yang, X., 2023.
High-resolution 3T to 7T MRI Synthesis with a Hybrid CNN-Transformer Model.
arXiv preprint arXiv:2311.15044.
Ekanayake, M., Pawar, K., Harandi, M., Egan, G., Chen, Z., 2024. McSTRA: A multi-branch
cascaded swin transformer for point spread function-guided robust MRI
reconstruction. Computers in Biology and Medicine 168, 107775.
Elmas, G., Dar, S.U., Korkmaz, Y., Ceyani, E., Susam, B., Ozbey, M., Avestimehr, S., Çukur, T.,
2022. Federated learning of generative image priors for MRI reconstruction. IEEE
Transactions on Medical Imaging.
Fang, J., Lin, H., Chen, X., Zeng, K., 2022. A hybrid network of cnn and transformer for
lightweight image super-resolution, in: Proceedings of the IEEE/CVF Conference on
Computer Vision and Pattern Recognition. pp. 1103–1112.
Feng, C.-M., Fu, H., Yuan, S., Xu, Y., 2021. Multi-contrast MRI super-resolution via a multi-
stage integration network, in: Medical Image Computing and Computer Assisted
Intervention–MICCAI 2021: 24th International Conference, Strasbourg, France,
September 27–October 1, 2021, Proceedings, Part VI 24. Springer, pp. 140–149.
Feng, C.-M., Yan, Y., Wang, S., Xu, Y., Shao, L., Fu, H., 2022. Specificity-preserving federated
learning for MR image reconstruction. IEEE Transactions on Medical Imaging.
Feng, R., Wu, Q., Feng, J., She, H., Liu, C., Zhang, Y., Wei, H., 2023. IMJENSE: scan-specific
implicit representation for joint coil sensitivity and image estimation in parallel
MRI. IEEE Transactions on Medical Imaging.

30
Feuerriegel, G.C., Goller, S.S., von Deuster, C., Sutter, R., 2023. Inflammatory Knee Synovitis:
Evaluation of an Accelerated FLAIR Sequence Compared With Standard Contrast-
Enhanced Imaging. Investigative Radiology 10–1097.
Gao, Z., Guo, Y., Zhang, J., Zeng, T., Yang, G., 2023. Hierarchical perception adversarial
learning framework for compressed sensing MRI. IEEE Transactions on Medical
Imaging.
Geng, C., Jiang, M., Fang, X., Li, Y., Jin, G., Chen, A., Liu, F., 2023. HFIST-Net: High-throughput
fast iterative shrinkage thresholding network for accelerating MR image
reconstruction. Computer Methods and Programs in Biomedicine 232, 107440.
Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A.,
Bengio, Y., 2020. Generative adversarial networks. Communications of the ACM 63,
139–144.
Güngör, A., Dar, S.U., Öztürk, Ş., Korkmaz, Y., Bedel, H.A., Elmas, G., Ozbey, M., Çukur, T.,
2023. Adaptive diffusion priors for accelerated MRI reconstruction. Medical Image
Analysis 88, 102872.
Guo, D., Zeng, G., Fu, H., Wang, Z., Yang, Y., Qu, X., 2023. A Joint Group Sparsity-based deep
learning for multi-contrast MRI reconstruction. Journal of Magnetic Resonance 346,
107354.
Guo, P., Mei, Y., Zhou, J., Jiang, S., Patel, V.M., 2023. ReconFormer: Accelerated MRI
reconstruction using recurrent transformer. IEEE transactions on medical imaging.
Guo, P., Valanarasu, J.M.J., Wang, P., Zhou, J., Jiang, S., Patel, V.M., 2021a. Over-and-under
complete convolutional rnn for mri reconstruction, in: Medical Image Computing
and Computer Assisted Intervention–MICCAI 2021: 24th International Conference,
Strasbourg, France, September 27–October 1, 2021, Proceedings, Part VI 24.
Springer, pp. 13–23.
Guo, P., Wang, P., Zhou, J., Jiang, S., Patel, V.M., 2021b. Multi-Institutional Collaborations for
Improving Deep Learning-Based Magnetic Resonance Image Reconstruction Using
Federated Learning, in: Proceedings of the IEEE/CVF Conference on Computer
Vision and Pattern Recognition (CVPR). pp. 2423–2432.
Haji-Valizadeh, H., Rahsepar, A.A., Collins, J.D., Bassett, E., Isakova, T., Block, T., Adluru, G.,
DiBella, E.V., Lee, D.C., Carr, J.C., others, 2018. Validation of highly accelerated real-
time cardiac cine MRI with radial k-space sampling and compressed sensing in
patients at 1.5 T and 3T. Magnetic resonance in medicine 79, 2745–2751.
Haldar, J.P., Hernando, D., Liang, Z.-P., 2010. Compressed-sensing MRI with random
encoding. IEEE transactions on Medical Imaging 30, 893–903.
Hammernik, K., Klatzer, T., Kobler, E., Recht, M.P., Sodickson, D.K., Pock, T., Knoll, F., 2018.
Learning a variational network for reconstruction of accelerated MRI data. Magnetic
resonance in medicine 79, 3055–3071.
Han, X., 2017. MR-based synthetic CT generation using a deep convolutional neural
network method. Medical physics 44, 1408–1419.
He, K., Zhang, X., Ren, S., Sun, J., 2016. Deep residual learning for image recognition, in:
Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition.
pp. 770–778.
Herrmann, J., Wessling, D., Nickel, D., Arberet, S., Almansour, H., Afat, C., Afat, S.,
Gassenmaier, S., Othman, A.E., 2023. Comprehensive clinical evaluation of a deep

31
 learning-accelerated, single-breath-hold abdominal HASTE at 1.5 T and 3 T.
Academic Radiology 30, 93–102.
Heydari, A., Ahmadi, A., Kim, T.H., Bilgic, B., 2024. Joint MAPLE: Accelerated joint T1 and T
2* T _2^ ∗ mapping with scan-specific self-supervised networks. Magnetic
Resonance in Medicine.
Hong, G.Q., Wei, Y.T., Morley, W.A., Wan, M., Mertens, A.J., Su, Y., Cheng, H.-L.M., 2023. Dual-
domain accelerated MRI reconstruction using transformers with learning-based
undersampling. Computerized Medical Imaging and Graphics 106, 102206.
Hricak, H., Abdel-Wahab, M., Atun, R., Lette, M.M., Paez, D., Brink, J.A., Donoso-Bach, L., Frija,
G., Hierath, M., Holmberg, O., others, 2021. Medical imaging and nuclear medicine: a
Lancet Oncology Commission. The Lancet Oncology 22, e136–e172.
Hu, C., Li, C., Wang, H., Liu, Q., Zheng, H., Wang, S., 2021. Self-supervised learning for mri
reconstruction with a parallel network training framework, in: Medical Image
Computing and Computer Assisted Intervention–MICCAI 2021: 24th International
Conference, Strasbourg, France, September 27–October 1, 2021, Proceedings, Part
VI 24. Springer, pp. 382–391.
Hu, M., Qian, J.Y., Pan, S., Li, Y., Qiu, R.L., Yang, X., 2024. Advancing medical imaging with
language models: featuring a spotlight on ChatGPT. Physics in Medicine and Biology.
Hu, Y., Xu, Y., Tian, Q., Chen, F., Shi, X., Moran, C.J., Daniel, B.L., Hargreaves, B.A., 2021. RUN-
UP: Accelerated multishot diffusion-weighted MRI reconstruction using an unrolled
network with U-Net as priors. Magnetic resonance in medicine 85, 709–720.
Huang, G., Liu, Z., Van Der Maaten, L., Weinberger, K.Q., 2017. Densely connected
convolutional networks, in: Proceedings of the IEEE Conference on Computer Vision
and Pattern Recognition. pp. 4700–4708.
Huang, J., Fang, Y., Wu, Y., Wu, H., Gao, Z., Li, Y., Del Ser, J., Xia, J., Yang, G., 2022. Swin
transformer for fast MRI. Neurocomputing 493, 281–304.
Hyun, C.M., Kim, H.P., Lee, S.M., Lee, S., Seo, J.K., 2018. Deep learning for undersampled MRI
reconstruction. Physics in Medicine & Biology 63, 135007.
Išgum, I., Benders, M.J., Avants, B., Cardoso, M.J., Counsell, S.J., Gomez, E.F., Gui, L., H\Huppi,
P.S., Kersbergen, K.J., Makropoulos, A., others, 2015. Evaluation of automatic
neonatal brain segmentation algorithms: the NeoBrainS12 challenge. Medical image
analysis 20, 135–151.
Isola, P., Zhu, J.-Y., Zhou, T., Efros, A.A., 2017. Image-to-image translation with conditional
adversarial networks, in: Proceedings of the IEEE Conference on Computer Vision
and Pattern Recognition. pp. 1125–1134.
Jiang, S., Eberhart, C.G., Lim, M., Heo, H.-Y., Zhang, Y., Blair, L., Wen, Z., Holdhoff, M., Lin, D.,
Huang, P., others, 2019. Identifying recurrent malignant glioma after treatment
using amide proton transfer-weighted MR imaging: a validation study with image-
guided stereotactic biopsy. Clinical Cancer Research 25, 552–561.
Jin, Z., Xiang, Q.-S., 2023. Improving accelerated MRI by deep learning with sparsified
complex data. Magnetic Resonance in Medicine 89, 1825–1838.
Johnson, P.M., Lin, D.J., Zbontar, J., Zitnick, C.L., Sriram, A., Muckley, M., Babb, J.S., Kline, M.,
Ciavarra, G., Alaia, E., others, 2023. Deep learning reconstruction enables
prospectively accelerated clinical knee MRI. Radiology 307, e220425.
Johnson, P.M., Tong, A., Donthireddy, A., Melamud, K., Petrocelli, R., Smereka, P., Qian, K.,
Keerthivasan, M.B., Chandarana, H., Knoll, F., 2022. Deep learning reconstruction
32
 enables highly accelerated biparametric MR imaging of the prostate. Journal of
Magnetic Resonance Imaging 56, 184–195.
Karras, T., Laine, S., Aila, T., 2019. A style-based generator architecture for generative
adversarial networks, in: Proceedings of the IEEE/CVF Conference on Computer
Vision and Pattern Recognition. pp. 4401–4410.
Khawaled, S., Freiman, M., 2024. NPB-REC: A non-parametric Bayesian deep-learning
approach for undersampled MRI reconstruction with uncertainty estimation.
Artificial Intelligence in Medicine 102798.
Kim, M., Chung, W., 2022. A cascade of preconditioned conjugate gradient networks for
accelerated magnetic resonance imaging. Computer Methods and Programs in
Biomedicine 225, 107090.
Kofler, A., Pali, M.-C., Schaeffter, T., Kolbitsch, C., 2023. Deep supervised dictionary learning
by algorithm unrolling—Application to fast 2D dynamic MR image reconstruction.
Medical Physics 50, 2939–2960.
Korkmaz, Y., Cukur, T., Patel, V.M., 2023. Self-supervised MRI reconstruction with unrolled
diffusion models, in: International Conference on Medical Image Computing and
Computer-Assisted Intervention. Springer, pp. 491–501.
Krizhevsky, A., Sutskever, I., Hinton, G.E., 2012. Imagenet classification with deep
convolutional neural networks. Advances in neural information processing systems
25.
Larkman, D.J., Nunes, R.G., 2007. Parallel magnetic resonance imaging. Phys. Med. Biol. 52,
R15–R55. https://doi.org/10.1088/0031-9155/52/7/R01
Lauterbur, P.C., 1973. Image Formation by Induced Local Interactions: Examples
Employing Nuclear Magnetic Resonance. Nature 242, 190–191.
https://doi.org/10.1038/242190a0
LeCun, Y., Bengio, Y., others, 1995. Convolutional networks for images, speech, and time
series. The handbook of brain theory and neural networks 3361, 1995.
Levac, B., Jalal, A., Tamir, J.I., 2023. Accelerated motion correction for MRI using score-
based generative models, in: 2023 IEEE 20th International Symposium on
Biomedical Imaging (ISBI). IEEE, pp. 1–5.
Levac, B.R., Arvinte, M., Tamir, J.I., 2023. Federated end-to-end unrolled models for
magnetic resonance image reconstruction. Bioengineering 10, 364.
Li, H., Xu, Z., Taylor, G., Studer, C., Goldstein, T., 2018. Visualizing the loss landscape of
neural nets. Advances in neural information processing systems 31.
Li, H., Yang, M., Kim, J.H., Zhang, C., Liu, R., Huang, P., Liang, D., Zhang, X., Li, X., Ying, L., 2023.
SuperMAP: Deep ultrafast MR relaxometry with joint spatiotemporal
undersampling. Magnetic resonance in medicine 89, 64–76.
Li, T., Sahu, A.K., Zaheer, M., Sanjabi, M., Talwalkar, A., Smith, V., 2020. Federated
optimization in heterogeneous networks. Proceedings of Machine learning and
systems 2, 429–450.
Li, Y., Joaquim, M.R., Pickup, S., Song, H.K., Zhou, R., Fan, Y., 2024. Learning ADC maps from
accelerated radial k-space diffusion-weighted MRI in mice using a deep CNN-
transformer model. Magnetic Resonance in Medicine 91, 105–117.
Liu, F., Kijowski, R., El Fakhri, G., Feng, L., 2021. Magnetic resonance parameter mapping
using model-guided self-supervised deep learning. Magnetic resonance in medicine
85, 3211–3226.
33
Liu, Q., Yang, Q., Cheng, H., Wang, S., Zhang, M., Liang, D., 2020. Highly undersampled
magnetic resonance imaging reconstruction using autoencoding priors. Magnetic
resonance in medicine 83, 322–336.
Liu, R., Zhang, Y., Cheng, S., Luo, Z., Fan, X., 2020. A deep framework assembling principled
modules for CS-MRI: unrolling perspective, convergence behaviors, and practical
modeling. IEEE Transactions on Medical Imaging 39, 4150–4163.
Liu, S., Li, H., Liu, Y., Cheng, G., Yang, G., Wang, H., Zheng, H., Liang, D., Zhu, Y., 2022. Highly
accelerated MR parametric mapping by undersampling the k-space and reducing the
contrast number simultaneously with deep learning. Physics in Medicine & Biology
67, 185004.
Liu, S., Liu, Y., Xu, X., Chen, R., Liang, D., Jin, Q., Liu, H., Chen, G., Zhu, Y., 2023. Accelerated
cardiac diffusion tensor imaging using deep neural network. Physics in Medicine &
Biology 68, 025008.
Liu, X., Wang, J., Sun, H., Chandra, S.S., Crozier, S., Liu, F., 2021. On the regularization of
feature fusion and mapping for fast MR multi-contrast imaging via iterative
networks. Magnetic resonance imaging 77, 159–168.
Liu, Y., Leong, A.T., Zhao, Y., Xiao, L., Mak, H.K., Tsang, A.C.O., Lau, G.K., Leung, G.K., Wu, E.X.,
2021. A low-cost and shielding-free ultra-low-field brain MRI scanner. Nature
communications 12, 7238.
Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B., 2021. Swin transformer:
Hierarchical vision transformer using shifted windows, in: Proceedings of the
IEEE/CVF International Conference on Computer Vision. pp. 10012–10022.
Lu, T., Zhang, X., Huang, Y., Guo, D., Huang, F., Xu, Q., Hu, Y., Ou-Yang, L., Lin, J., Yan, Z.,
others, 2020. pFISTA-SENSE-ResNet for parallel MRI reconstruction. Journal of
Magnetic Resonance 318, 106790.
Lyu, J., Tian, Y., Cai, Q., Wang, C., Qin, J., 2023. Adaptive channel-modulated personalized
federated learning for magnetic resonance image reconstruction. Computers in
Biology and Medicine 165, 107330.
Lyu, Q., Shan, H., Steber, C., Helis, C., Whitlow, C., Chan, M., Wang, G., 2020. Multi-contrast
super-resolution MRI through a progressive network. IEEE transactions on medical
imaging 39, 2738–2749.
Maclaren, J., Herbst, M., Speck, O., Zaitsev, M., 2013. Prospective motion correction in brain
imaging: a review. Magnetic resonance in medicine 69, 621–636.
Mardani, M., Gong, E., Cheng, J.Y., Vasanawala, S.S., Zaharchuk, G., Xing, L., Pauly, J.M., 2018.
Deep generative adversarial neural networks for compressive sensing MRI. IEEE
transactions on medical imaging 38, 167–179.
McMahan, B., Moore, E., Ramage, D., Hampson, S., y Arcas, B.A., 2017. Communication-
efficient learning of deep networks from decentralized data, in: Artificial Intelligence
and Statistics. PMLR, pp. 1273–1282.
Mendrik, A.M., Vincken, K.L., Kuijf, H.J., Breeuwer, M., Bouvy, W.H., De Bresser, J., Alansary,
A., De Bruijne, M., Carass, A., El-Baz, A., others, 2015. MRBrainS challenge: online
evaluation framework for brain image segmentation in 3T MRI scans.
Computational intelligence and neuroscience 2015, 1–1.
Meng, Z., Guo, R., Li, Yudu, Guan, Y., Wang, T., Zhao, Y., Sutton, B., Li, Yao, Liang, Z.-P., 2021.
Accelerating T2 mapping of the brain by integrating deep learning priors with low-
rank and sparse modeling. Magnetic Resonance in Medicine 85, 1455–1467.
34
Menze, B.H., Jakab, A., Bauer, S., Kalpathy-Cramer, J., Farahani, K., Kirby, J., Burren, Y., Porz,
N., Slotboom, J., Wiest, R., others, 2014. The multimodal brain tumor image
segmentation benchmark (BRATS). IEEE transactions on medical imaging 34, 1993–
2024.
Mirza, M., Osindero, S., 2014. Conditional generative adversarial nets. arXiv preprint
arXiv:1411.1784.
Murali, S., Ding, H., Adedeji, F., Qin, C., Obungoloch, J., Asllani, I., Anazodo, U., Ntusi, N.A.,
Mammen, R., Niendorf, T., others, 2023. Bringing MRI to low-and middle-income
countries: directions, challenges and potential solutions. NMR in Biomedicine
e4992.
Murugesan, B., Ramanarayanan, S., Vijayarangan, S., Ram, K., Jagannathan, N.R.,
Sivaprakasam, M., 2021. A deep cascade of ensemble of dual domain networks with
gradient-based T1 assistance and perceptual refinement for fast MRI reconstruction.
Computerized Medical Imaging and Graphics 91, 101942.
Narnhofer, D., Hammernik, K., Knoll, F., Pock, T., 2019. Inverse GANs for accelerated MRI
reconstruction, in: Wavelets and Sparsity XVIII. SPIE, pp. 381–392.
Pan, S., Abouei, E., Wynne, J., Chang, C.-W., Wang, T., Qiu, R.L., Li, Y., Peng, J., Roper, J., Patel,
P., others, 2023a. Synthetic CT generation from MRI using 3D transformer-based
denoising diffusion model. Medical Physics.
Pan, S., Wang, T., Qiu, R.L., Axente, M., Chang, C.-W., Peng, J., Patel, A.B., Shelton, J., Patel, S.A.,
Roper, J., others, 2023b. 2D medical image synthesis using transformer-based
denoising diffusion probabilistic model. Physics in Medicine & Biology 68, 105004.
Pramanik, A., Bhave, S., Sajib, S., Sharma, S.D., Jacob, M., 2023. Adapting model-based deep
learning to multiple acquisition conditions: Ada-MoDL. Magnetic resonance in
medicine 90, 2033–2051.
Qiao, X., Huang, Y., Li, W., 2023. MEDL-Net: A model-based neural network for MRI
reconstruction with enhanced deep learned regularizers. Magnetic Resonance in
Medicine 89, 2062–2075.
Qin, C., Duan, J., Hammernik, K., Schlemper, J., Küstner, T., Botnar, R., Prieto, C., Price, A.N.,
Hajnal, J.V., Rueckert, D., 2021. Complementary time-frequency domain networks
for dynamic parallel MR image reconstruction. Magnetic Resonance in Medicine 86,
3274–3291.
Qin, C., Schlemper, J., Caballero, J., Price, A.N., Hajnal, J.V., Rueckert, D., 2018. Convolutional
recurrent neural networks for dynamic MR image reconstruction. IEEE transactions
on medical imaging 38, 280–290.
Qu, B., Zhang, J., Kang, T., Lin, J., Lin, M., She, H., Wu, Q., Wang, M., Zheng, G., 2024. Radial
magnetic resonance image reconstruction with a deep unrolled projected fast
iterative soft-thresholding network. Computers in Biology and Medicine 168,
107707.
Quan, T.M., Nguyen-Duc, T., Jeong, W.-K., 2018. Compressed sensing MRI reconstruction
using a generative adversarial network with a cyclic loss. IEEE transactions on
medical imaging 37, 1488–1497.
Radmanesh, A., Muckley, M.J., Murrell, T., Lindsey, E., Sriram, A., Knoll, F., Sodickson, D.K.,
Lui, Y.W., 2022. Exploring the acceleration limits of deep learning variational
network–based two-dimensional brain MRI. Radiology: Artificial Intelligence 4,
e210313.
35
Ravishankar, S., Bresler, Y., 2010. MR image reconstruction from highly undersampled k-
space data by dictionary learning. IEEE transactions on medical imaging 30, 1028–
1041.
Rohlfing, T., Zahr, N.M., Sullivan, E.V., Pfefferbaum, A., 2010. The SRI24 multichannel atlas
of normal adult human brain structure. Human brain mapping 31, 798–819.
Ronneberger, O., Fischer, P., Brox, T., 2015. U-net: Convolutional networks for biomedical
image segmentation, in: Medical Image Computing and Computer-Assisted
Intervention–MICCAI 2015: 18th International Conference, Munich, Germany,
October 5-9, 2015, Proceedings, Part III 18. Springer, pp. 234–241.
Ryu, K., Lee, J.-H., Nam, Y., Gho, S.-M., Kim, H.-S., Kim, D.-H., 2021. Accelerated multicontrast
reconstruction for synthetic MRI using joint parallel imaging and variable splitting
networks. Medical physics 48, 2939–2950.
Safari, M., Fatemi, A., Archambault, L., 2023a. MedFusionGAN: multimodal medical image
fusion using an unsupervised deep generative adversarial network. BMC Medical
Imaging 23, 203.
Safari, M., Yang, X., Fatemi, A., 2024. MRI data consistency guided conditional diffusion
probabilistic model for MR imaging acceleration, in: Medical Imaging 2024: Clinical
and Biomedical Imaging. SPIE, pp. 202–205.
Safari, M., Yang, X., Fatemi, A., Archambault, L., 2023b. MRI motion artifact reduction using
a conditional diffusion probabilistic model (MAR-CDPM). Medical Physics.
Sandino, C.M., Lai, P., Vasanawala, S.S., Cheng, J.Y., 2021. Accelerating cardiac cine MRI using
a deep learning-based ESPIRiT reconstruction. Magnetic Resonance in Medicine 85,
152–167.
Schlemper, J., Caballero, J., Hajnal, J.V., Price, A., Rueckert, D., 2017. A deep cascade of
convolutional neural networks for MR image reconstruction, in: Information
Processing in Medical Imaging: 25th International Conference, IPMI 2017, Boone,
NC, USA, June 25-30, 2017, Proceedings 25. Springer, pp. 647–658.
Seo, S., Luu, H.M., Choi, S.H., Park, S.-H., 2022. Simultaneously optimizing sampling pattern
for joint acceleration of multi-contrast MRI using model-based deep learning.
Medical Physics 49, 5964–5980.
Shaul, R., David, I., Shitrit, O., Raviv, T.R., 2020. Subsampled brain MRI reconstruction by
generative adversarial neural networks. Medical Image Analysis 65, 101747.
Shitrit, O., Riklin Raviv, T., 2017. Accelerated magnetic resonance imaging by adversarial
neural network, in: Deep Learning in Medical Image Analysis and Multimodal
Learning for Clinical Decision Support: Third International Workshop, DLMIA 2017,
and 7th International Workshop, ML-CDS 2017, Held in Conjunction with MICCAI
2017, Québec City, QC, Canada, September 14, Proceedings 3. Springer, pp. 30–38.
Slavkova, K.P., DiCarlo, J.C., Wadhwa, V., Kumar, S., Wu, C., Virostko, J., Yankeelov, T.E.,
Tamir, J.I., 2023. An untrained deep learning method for reconstructing dynamic MR
images from accelerated model-based data. Magnetic resonance in medicine 89,
1617–1633.
Slipsager, J.M., Glimberg, S.L., Søgaard, J., Paulsen, R.R., Johannesen, H.H., Martens, P.C., Seth,
A., Marner, L., Henriksen, O.M., Olesen, O.V., others, 2020. Quantifying the financial
savings of motion correction in brain MRI: a model-based estimate of the costs
arising from patient head motion and potential savings from implementation of
motion correction. Journal of Magnetic Resonance Imaging 52, 731–738.
36
Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S., 2015. Deep unsupervised
learning using nonequilibrium thermodynamics, in: International Conference on
Machine Learning. PMLR, pp. 2256–2265.
Sun, K., Wang, Q., Shen, D., 2023. Joint Cross-Attention Network with Deep Modality Prior
for Fast MRI Reconstruction. IEEE Transactions on Medical Imaging.
Sun, L., Fan, Z., Fu, X., Huang, Y., Ding, X., Paisley, J., 2019. A deep information sharing
network for multi-contrast compressed sensing MRI reconstruction. IEEE
Transactions on Image Processing 28, 6141–6153.
Sun, L., Huang, Y., Cai, C., Ding, X., 2017. Compressed sensing MRI using total variation
regularization with K-space decomposition, in: 2017 IEEE International Conference
on Image Processing (ICIP). IEEE, pp. 3061–3065.
Terpstra, M.L., Maspero, M., Verhoeff, J.J., van den Berg, C.A., 2023. Accelerated respiratory-
resolved 4D-MRI with separable spatio-temporal neural networks. Medical physics
50, 5331–5342.
Uecker, M., Lai, P., Murphy, M.J., Virtue, P., Elad, M., Pauly, J.M., Vasanawala, S.S., Lustig, M.,
2014. ESPIRiT—an eigenvalue approach to autocalibrating parallel MRI: where
SENSE meets GRAPPA. Magnetic resonance in medicine 71, 990–1001.
Van Essen, D.C., Smith, S.M., Barch, D.M., Behrens, T.E., Yacoub, E., Ugurbil, K., Consortium,
W.-M.H., others, 2013. The WU-Minn human connectome project: an overview.
Neuroimage 80, 62–79.
Wang, B., Lian, Y., Xiong, X., Zhou, H., Liu, Z., Zhou, X., 2024. DCT-net: Dual-domain cross-
fusion transformer network for MRI reconstruction. Magnetic Resonance Imaging
107, 69–79.
Wang, J., Awad, M., Zhou, R., Wang, Z., Wang, X., Feng, X., Yang, Y., Meyer, C., Kramer, C.M.,
Salerno, M., 2024. High-resolution spiral real-time cardiac cine imaging with deep
learning-based rapid image reconstruction and quantification. NMR in Biomedicine
37, e5051.
Wang, K., Tamir, J.I., De Goyeneche, A., Wollner, U., Brada, R., Yu, S.X., Lustig, M., 2022. High
fidelity deep learning-based MRI reconstruction with instance-wise discriminative
feature matching loss. Magnetic Resonance in Medicine 88, 476–491.
Wang, S., Cheng, H., Ying, L., Xiao, T., Ke, Z., Zheng, H., Liang, D., 2020. DeepcomplexMRI:
Exploiting deep residual network for fast parallel MR imaging with complex
convolution. Magnetic resonance imaging 68, 136–147.
Wang, Z., Fang, H., Qian, C., Shi, B., Bao, L., Zhu, L., Zhou, J., Wei, W., Lin, J., Guo, D., others,
2024. A faithful deep sensitivity estimation for accelerated magnetic resonance
imaging. IEEE Journal of Biomedical and Health Informatics.
Wang, Z., Qian, C., Guo, D., Sun, H., Li, R., Zhao, B., Qu, X., 2022. One-dimensional deep low-
rank and sparse network for accelerated MRI. IEEE Transactions on Medical
Imaging 42, 79–90.
Wang, Z., She, H., Zhang, Y., Du, Y.P., 2023. Parallel non-Cartesian spatial-temporal
dictionary learning neural networks (stDLNN) for accelerating 4D-MRI. Medical
image analysis 84, 102701.
Wen, J., Ahmad, R., Schniter, P., 2023. A conditional normalizing flow for accelerated multi-
coil MR imaging. Proceedings of machine learning research 202, 36926.
Wu, Y., Ma, Y., Liu, J., Du, J., Xing, L., 2019. Self-attention convolutional neural network for
improved MR image reconstruction. Information sciences 490, 317–328.
37
Wu, Z., Liao, W., Yan, C., Zhao, M., Liu, G., Ma, N., Li, X., 2023. Deep learning based MRI
reconstruction with transformer. Computer Methods and Programs in Biomedicine
233, 107452.
Xiang, L., Chen, Y., Chang, W., Zhan, Y., Lin, W., Wang, Q., Shen, D., 2018. Ultra-fast t2-
weighted mr reconstruction using complementary t1-weighted information, in:
Medical Image Computing and Computer Assisted Intervention–MICCAI 2018: 21st
International Conference, Granada, Spain, September 16-20, 2018, Proceedings, Part
I. Springer, pp. 215–223.
Xie, H., Lei, Y., Wang, T., Roper, J., Dhabaan, A.H., Bradley, J.D., Liu, T., Mao, H., Yang, X., 2022.
Synthesizing high-resolution magnetic resonance imaging using parallel cycle-
consistent generative adversarial networks for fast magnetic resonance imaging.
Medical Physics 49, 357–369.
Xie, Y., Li, Q., 2022. A Review of Deep Learning Methods for Compressed Sensing Image
Reconstruction and Its Medical Applications. Electronics 11, 586.
https://doi.org/10.3390/electronics11040586
Xu, J., Zu, T., Hsu, Y.-C., Wang, X., Chan, K.W., Zhang, Y., 2024. Accelerating CEST imaging
using a model-based deep neural network with synthetic training data. Magnetic
Resonance in Medicine 91, 583–599.
Xue, Y., Yu, J., Kang, H.S., Englander, S., Rosen, M.A., Song, H.K., 2012. Automatic coil
selection for streak artifact reduction in radial MRI. Magnetic Resonance in Med 67,
470–476. https://doi.org/10.1002/mrm.23023
Yaman, B., Gu, H., Hosseini, S.A.H., Demirel, O.B., Moeller, S., Ellermann, J., Uğurbil, K.,
Akçakaya, M., 2022. Multi-mask self-supervised learning for physics-guided neural
networks in highly accelerated magnetic resonance imaging. NMR in Biomedicine
35, e4798.
Yaman, B., Hosseini, S.A.H., Moeller, S., Ellermann, J., Uğurbil, K., Akçakaya, M., 2020. Self-
supervised learning of physics-guided reconstruction neural networks without fully
sampled reference data. Magnetic resonance in medicine 84, 3172–3191.
Yan, Y., Yang, T., Zhao, X., Jiao, C., Yang, A., Miao, J., 2023. DC-SiamNet: Deep contrastive
Siamese network for self-supervised MRI reconstruction. Computers in Biology and
Medicine 167, 107619.
Yang, G., Yu, S., Dong, H., Slabaugh, G., Dragotti, P.L., Ye, X., Liu, F., Arridge, S., Keegan, J., Guo,
Y., others, 2017. DAGAN: deep de-aliasing generative adversarial networks for fast
compressed sensing MRI reconstruction. IEEE transactions on medical imaging 37,
1310–1321.
Yang, Q., Liu, Y., Chen, T., Tong, Y., 2019. Federated machine learning: Concept and
applications. ACM Transactions on Intelligent Systems and Technology (TIST) 10, 1–
19.
Yang, Y., Sun, J., Li, H., Xu, Z., 2016. Deep ADMM-Net for compressive sensing MRI, in:
Proceedings of the 30th International Conference on Neural Information Processing
Systems. pp. 10–18.
Yiasemis, G., Sánchez, C.I., Sonke, J.-J., Teuwen, J., 2024. On retrospective k-space
subsampling schemes for deep MRI reconstruction. Magnetic Resonance Imaging
107, 33–46.

38
Yoon, M.A., Gold, G.E., Chaudhari, A.S., 2023. Accelerated Musculoskeletal Magnetic
Resonance Imaging. Magnetic Resonance Imaging jmri.29205.
https://doi.org/10.1002/jmri.29205
Zbontar, J., Knoll, F., Sriram, A., Murrell, T., Huang, Z., Muckley, M.J., Defazio, A., Stern, R.,
Johnson, P., Bruno, M., others, 2018. fastMRI: An open dataset and benchmarks for
accelerated MRI. arXiv preprint arXiv:1811.08839.
Zhao, X., Yang, T., Li, B., Zhang, X., 2023. SwinGAN: A dual-domain Swin Transformer-based
generative adversarial network for MRI reconstruction. Computers in Biology and
Medicine 153, 106513.
Zhou, B., Schlemper, J., Dey, N., Salehi, S.S.M., Sheth, K., Liu, C., Duncan, J.S., Sofka, M., 2022.
Dual-domain self-supervised learning for accelerated non-Cartesian MRI
reconstruction. Medical Image Analysis 81, 102538.
Zhou, Z., Han, F., Ghodrati, V., Gao, Y., Yin, W., Yang, Y., Hu, P., 2019. Parallel imaging and
convolutional neural network combined fast MR image reconstruction: Applications
in low-latency accelerated real-time imaging. Medical physics 46, 3399–3413.
Zhu, J.-Y., Park, T., Isola, P., Efros, A.A., 2017. Unpaired image-to-image translation using
cycle-consistent adversarial networks, in: Proceedings of the IEEE International
Conference on Computer Vision. pp. 2223–2232.
Zufiria, B., Qiu, S., Yan, K., Zhao, R., Wang, R., She, H., Zhang, C., Sun, B., Herman, P., Du, Y.,
others, 2022. A feature-based convolutional neural network for reconstruction of
interventional MRI. NMR in Biomedicine 35, e4231.

39