A Transfer-Learning Approach for Accelerated MRI using Deep Neural Networks

Salman Ul Hassan Dar1,2, Muzaffer Özbey1,2, Ahmet Burak Çatlı1,2, Tolga Çukur1,2,3

1Department of Electrical and Electronics Engineering, Bilkent University, Ankara, Turkey

2National Magnetic Resonance Research Center (UMRAM), Bilkent University, Ankara,
Turkey
3Neuroscience Program, Sabuncu Brain Research Center, Bilkent University, Ankara, Turkey

Running title: A transfer-learning approach for accelerated MRI using deep neural networks

Address correspondence to:

Tolga Çukur
Department of Electrical and Electronics Engineering, Room 304
Bilkent University
Ankara, TR-06800, Turkey
TEL: +90 (312) 290-1164
E-MAIL: [email protected]

This work was supported in part by a Marie Curie Actions Career Integration Grant (PCIG13-
GA- 2013-618101), by a European Molecular Biology Organization Installation Grant (IG
3028), by a TUBA GEBIP fellowship, and by a BAGEP fellowship awarded to T. Çukur. We
also gratefully acknowledge the support of NVIDIA Corporation with the donation of the Titan
X Pascal GPU used for this research.

To be submitted to Magnetic Resonance in Medicine.

 Abstract

Purpose: Neural networks have received recent interest for reconstruction of undersampled
MR acquisitions. Ideally network performance should be optimized by drawing the training and
testing data from the same domain. In practice, however, large datasets comprising hundreds of
subjects scanned under a common protocol are rare. The goal of this study is to introduce a
transfer-learning approach to address the problem of data scarcity in training deep networks for
accelerated MRI.

Methods: Neural networks were trained on thousands of samples from public datasets of either
natural images or brain MR images. The networks were then fine-tuned using only few tens of
brain MR images in a distinct testing domain. Domain-transferred networks were compared to
networks trained directly in the testing domain. Network performance was evaluated for
varying acceleration factors (2-10), number of training samples (0.5-4k) and number of fine-
tuning samples (0-100).

Results: The proposed approach achieves successful domain transfer between MR images
acquired with different contrasts (T1- and T2-weighted images), and between natural and MR
images (ImageNet and T1- or T2-weighted images). Networks obtained via transfer-learning
using only tens of images in the testing domain achieve nearly identical performance to
networks trained directly in the testing domain using thousands of images.

Conclusion: The proposed approach might facilitate the use of neural networks for MRI
reconstruction without the need for collection of extensive imaging datasets.
 Introduction

The unparalleled soft-tissue contrast in MRI has rendered it a preferred modality in many
diagnostic applications, but long scan durations limit its clinical use. Acquisitions can be
accelerated by undersampling in k-space, and a tailored reconstruction can be used to recover
unacquired data. Because MR images are inherently compressible, a popular framework for
accelerated MRI has been compressive sensing (CS) (1,2). CS has offered improvements in
scan efficiency in many applications including structural (2), angiographic (3), functional (4),
diffusion (5), and parametric imaging (6). Yet the CS framework is not without limitation. First,
CS involves nonlinear optimization algorithms that scale poorly with growing data size and
hamper clinical workflow. Second, CS commonly assumes that MRI data are sparse in fixed
transform domains, such as finite differences or wavelet transforms. Recent studies highlight
the need for learning the transform domains specific to each dataset to optimize performance
(7). Lastly, CS requires careful parameter tuning (e.g., for regularization) for optimal
performance. While several approaches were proposed for data-driven parameter tuning (8,9),
these methods can induce further computational burden.

Neural network (NN) architectures that reconstruct images from undersampled data have
recently been proposed to address the abovementioned limitations. Improved image quality
over traditional CS has readily been demonstrated for several applications including
angiographic (10), cardiac (11–13), brain (14–34), abdominal (35–37), and musculoskeletal
imaging (38–42). The common approach is to train a network off-line using a relatively large
set of fully-sampled MRI data, and then use it for on-line reconstruction of undersampled data.
Reconstructions can be achieved in several hundred milliseconds, significantly reducing
computational burden (39,43). The NN framework also alleviates the need for adhoc selection
of transform domains. For example, a recent study used a cascade of CNNs to recover images
directly from zero-filled Fourier reconstructions of undersampled data (11,16,41). The trained
CNN layers reflect suitable transforms for image reconstruction. The NN framework introduces
more tunable hyperparameters (e.g., number of layers, units, activation functions) than would
be required in CS. However, previous studies demonstrate that hyperparameters optimized
during the training phase generally perform well in the testing phase (43). Taken together, these
advantages render the NN framework a promising avenue for accelerated MRI.

A common strategy to enhance network performance is to boost model complexity by

increasing the number of layers and units in the architecture. A large set of training data must
then be used to reliably learn the numerous model parameters (44). Previous studies either used
an extensive database of MR images comprising several tens to hundreds of subjects (12,18,45),
or data augmentation procedures to artificially expand the size of training data (11,12). For
instance, an early study performed training on T1-weighted brain images from nearly 500
subjects in the human connectome project (HCP) database, and testing on T2-weighted images
(18). Yet, it remains unclear how well a network trained on images acquired with a specific
type of tissue contrast generalize to images acquired with different contrasts. Furthermore, for
optimal reconstruction performance the network must be trained on images acquired with the
same scan protocol that it later will be tested on. However, large databases such as those
provided by the HCP may not be readily available in many applications, potentially rendering
NN-based reconstructions suboptimal.

In this study, we propose a transfer-learning approach to address the problem of data scarcity
in network training for accelerated MRI. In transfer-learning, network training is performed in
some domain where large datasets are available, and knowledge captured by the trained network
is then transferred to a different domain where data are scarce (46,47). Domain transfer was
previously used to suppress coherent aliasing artifacts in projection reconstruction acquisitions
(30), to perform non-Cartesian to Cartesian interpolation in k-space (18), and to assess the
robustness of network reconstructions to variations in SNR and undersampling patterns (42).
In contrast, we employ transfer-learning to enhance NN-based reconstructions of randomly
undersampled acquisitions in the testing domain. A deep CNN architecture with multiple
subnetworks is taken as a model network (11). For reconstruction of multi-coil data, calibration
consistency, data consistency and CNN blocks are incorporated to synthesize missing samples.
In the training domain using several thousand images, each subnetwork is trained sequentially
to reconstruct reference images from zero-filled reconstructions of undersampled data. The full
network is then fine-tuned end-to-end in the testing domain using few tens of images.

To demonstrate the proposed approach, comprehensive evaluations were performed across a

broad range of acceleration factors (R=2-10) on T1- and T2-weighted brain images. Separate
network models were learnt for domain transfer between MR images acquired with different
contrasts (T1- and T2-weighted), and between natural and MR images (ImageNet and T1- or T2-
weighted). Domain-transferred networks were quantitatively compared against networks
trained in the testing domain, and against conventional CS reconstructions. We find that
domain-transferred networks fine-tuned with tens of images achieve nearly identical
performance to networks trained directly in the testing domain using thousands of images, and
that both networks outperform CS methods.

Note – A preliminary version of this work was presented at the 26th Annual Meeting of ISMRM
under the title “Transfer learning for reconstruction of accelerated MRI acquisitions via neural
networks” (48).
 Methods

MRI Reconstruction via Compressed Sensing (CS)

Single-coil data. In accelerated MRI, an undersampled acquisition is followed by a

reconstruction to recover missing k-space samples. This recovery can be formulated as a linear
inverse problem:

𝐹𝑢 𝑥 = 𝑦𝑢 (1)

where 𝑥 denotes the image to be reconstructed, 𝐹𝑢 is the partial Fourier transform operator at the
sampled k-space locations, and 𝑦𝑢 denotes acquired k-space data. Since Eq. 1 is
underdetermined, additional prior information is typically incorporated in the form of a
regularization term:

𝑥𝑟𝑒𝑐 = min‖𝐹𝑢 𝑥 − 𝑦𝑢 ‖2 + 𝑅(𝑥) (2)

𝑥

Here, the first term enforces consistency between acquired and reconstructed data, whereas 𝑅(𝑥)
enforces prior information to improve reconstruction performance. In CS, 𝑅(𝑥) typically
corresponds to L1-norm of the image in a known transform domain (e.g., wavelet transform or
finite differences transform).

The solution of Eq. 2 involves non-linear optimization algorithms that are often
computationally complex. This reduces clinical feasibility as reconstructions times become
prohibitive with increasing size of data. Furthermore, assuming ad hoc selection of fixed
transform domains leads to suboptimal reconstructions in many applications (7). Lastly, it is
often challenging to find a set of reconstruction parameters that work optimally across subjects
(49).

Multi-coil data. For reconstruction of multi-coil data, a hybrid parallel imaging/compressed

sensing approach is commonly used, such as SPIRiT (iTerative Self-consistent Parallel Imaging
Reconstruction). In SPIRiT (50), the recovery problem in Eq. 2 can be reformulated as:

𝑥𝑟𝑒𝑐 = min‖𝐹𝑢 𝑥 − 𝑦𝑢 ‖2 + ‖(𝐺 − 𝐼)𝐹𝑥‖2 + 𝑅(𝑥) (3)

𝑥

where 𝑥 denotes multi-coil images to be reconstructed, 𝑦𝑢 denotes acquired multi-coil k-space

data, 𝐹 is the forward Fourier transform operator and 𝐺 denotes the interpolation operator that
synthesizes unacquired samples in terms of acquired samples across neighboring k-space and
coils. To enforce sparsity, 𝑅(𝑥) can be selected as the L1-norm of wavelet coefficients. One
efficient way to solve Eq. 3 is via the projection onto convex sets (POCS) algorithm (51). POCS
alternates among a calibration-consistency (CC) projection that applies G, a sparsity projection
that enforces sparsity in the transform domain, and a data-consistency (DC) projection.

MRI Reconstruction via Neural Networks (NN)

Single-coil data. In the NN framework, a network architecture is used for reconstruction instead
of explicit transform-domain constraints. Network training is performed via a supervised
learning procedure, with the aim to find the set of network parameters that yield accurate
reconstructions undersampled acquisitions. This procedure is performed on a large set of
training data (with 𝑁𝑡𝑟𝑎𝑖𝑛 samples), where fully-sampled reference acquisitions are
retrospectively undersampled. Network training typically amounts to minimizing the following
loss function (29):
𝑁𝑡𝑟𝑎𝑖𝑛
1
min ∑ ‖𝐶(𝑥𝑢𝑛 ; 𝜃) − 𝑥𝑟𝑒𝑓𝑛 ‖2 (4)
𝜃 𝑁𝑡𝑟𝑎𝑖𝑛
𝑛=1

where 𝑥𝑢𝑛 represents the Fourier reconstruction of nth undersampled acquisition, 𝑥𝑟𝑒𝑓𝑛 represents
the respective Fourier reconstruction of the fully-sampled acquisition, 𝐶(𝑥𝑢𝑛 ; 𝜃) denotes the
output of the network given the input image 𝑥𝑢𝑛 and the network parameters 𝜃 . To reduce
sensitivity to outliers, here we minimized a hybrid loss that includes both mean-squared error
and mean-absolute error terms. To minimize over-fitting, we further added an L2-regularization
term on the network parameters. Therefore, neural network training was performed with the
following loss function:
𝑁𝑡𝑟𝑎𝑖𝑛 𝑁𝑡𝑟𝑎𝑖𝑛
1 1
min ∑ ‖𝐶(𝑥𝑢𝑛 ; 𝜃) − 𝑥𝑟𝑒𝑓𝑛 ‖2 + ∑ ‖𝐶(𝑥𝑢𝑛 ; 𝜃) − 𝑥𝑟𝑒𝑓𝑛 ‖1 + 𝛾𝛷 ‖𝜃‖2 (5)
𝜃 𝑁𝑡𝑟𝑎𝑖𝑛 𝑁𝑡𝑟𝑎𝑖𝑛
𝑛=1 𝑛=1

where 𝛾𝛷 is the regularization parameter for network parameters.

A network trained on a sufficiently large set of training examples can then be used to reconstruct
an undersampled acquisition from an independent test dataset. This reconstruction can be
achieved by reformulating the problem in Eq. 2 (29):

𝑥𝑟𝑒𝑐 = min 𝜆‖𝐹𝑢 𝑥 − 𝑦𝑢 ‖2 + ‖𝐶(𝑥𝑢 ; 𝜃 ∗ ) − 𝑥‖2 (6)

𝑥

where 𝐶(𝑥𝑢 ; 𝜃 ∗ ) is the output of the trained network with optimized parameters 𝜃 ∗ . Note that the
problem in Eq. 6 has the following closed-form solution (29):

[𝐹𝐶(𝑥𝑢 ; 𝜃 ∗ )](𝑘) + 𝜆𝑦𝑢 (𝑘)

, 𝑖𝑓 𝑘 𝜖 𝛺
𝑦𝑟𝑒𝑐 (𝑘) = { 1+𝜆
[𝐹𝐶(𝑥𝑢 ; 𝜃 ∗ )](𝑘), 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒 (7)

𝑥𝑟𝑒𝑐 = 𝐹 −1 𝑦𝑟𝑒𝑐

where 𝑘 denotes k-space location, 𝛺 represents the set of acquired k-space locations, 𝐹 and 𝐹 −1
are the forward and backward Fourier transform operators, and 𝑥𝑟𝑒𝑐 is the reconstructed image.
The solution outlined in Eq. 7 performs two separate projections during reconstruction. The
first projection calculates the output of the trained neural network 𝐶(𝑥𝑢 ; 𝜃 ∗ ) given the input
image 𝑥𝑢 , the Fourier reconstruction of undersampled data. The second projection enforces data
consistency. The parameter 𝜆 in Eq. 7 controls the relative weighing between data samples that
are originally acquired and those that are recovered by the network. Here we used 𝜆 = ∞ to
enforce data consistency strictly. The projection outlined in Eq. 7 can be compactly expressed
as (11):

𝜆
𝑓𝐷𝐶 {𝐶(𝑥𝑢 ; 𝜃 ∗ )} = 𝐹 −1 𝛬𝐹𝐶(𝑥𝑢 ; 𝜃 ∗ ) + 𝐹 −1 𝑥𝑢 (8)
1+𝜆

where 𝛬 is a diagonal matrix:

 1
𝛬𝑘𝑘 = { 1 + 𝜆 , 𝑖𝑓 𝑘 𝜖 𝛺 (9)
1, 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒

Conventional optimization algorithms for CS run iteratively to progressively minimize the loss
function. A similar approach can also be adopted for NN-based reconstructions (11,16,41).
Here, we cascaded several subnetworks in series with DC projections interleaved between
consecutive subnetworks (11). In this architecture, the input 𝑥𝑖𝑝 to the 𝑝𝑡ℎ subnetwork was
formed as:

𝑥𝑢𝑛 , 𝑖𝑓 𝑝 = 1
𝑥𝑖𝑝 = { (10)
𝑓𝐷𝐶 {𝐶𝑝−1 (𝑓𝐷𝐶 {𝐶𝑝−2 (𝑓𝐷𝐶 … . 𝐶1 (𝑥𝑢𝑛 ; 𝜃1∗ )}; 𝜃𝑝−1
∗
)}, 𝑖𝑓 𝑝 > 1

where 𝜃𝑝∗ denotes the parameters of the 𝑝𝑡ℎ subnetwork. Starting with the initial network with 𝑝
= 1, each subnetwork was trained sequentially by solving the following optimization problem:
𝑁𝑡𝑟𝑎𝑖𝑛 𝑁𝑡𝑟𝑎𝑖𝑛
1 1 2
min ∑ ‖𝐶(𝑥𝑖𝑝 ; 𝜃𝑝 ) − 𝑥𝑟𝑒𝑓𝑛 ‖2 + ∑ ‖𝐶(𝑥𝑖𝑝 ; 𝜃𝑝 ) − 𝑥𝑟𝑒𝑓𝑛 ‖1 + 𝛾𝛷 ‖𝜃𝑝 ‖ (11)
𝜃𝑝 𝑁𝑡𝑟𝑎𝑖𝑛 𝑁𝑡𝑟𝑎𝑖𝑛
𝑛=1 𝑛=1

While training the 𝑝𝑡ℎ subnetwork, the parameters of preceding networks and thus the input 𝑥𝑖𝑝
are assumed to be fixed.

Multi-coil data. Similar to SPIRiT, for multi-coil reconstructions, here we reformulate Eq. 3 as:

𝑥𝑟𝑒𝑐 = min‖𝐹𝑢 𝑥 − 𝑦𝑢 ‖2 + ‖(𝐺 − 𝐼)𝐹𝑥‖2 + ‖𝐶(𝐴∗ 𝑥𝑢 ; 𝜃 ∗ ) − 𝐴∗ 𝑥‖2 (12)

𝑥

where 𝑥 denotes the multi-coil images to be reconstructed, 𝐴 and 𝐴∗ denote coil-sensitivity

profiles and adjoints, and G denotes the interpolation operator in SPIRiT as in Eq. 3. The
network C has been trained to recover fully-sampled coil-combined images given undersampled
coil-combined images as outlined in Eq. 5. The trained network regularizes the reconstruction
in Eq. 12 given undersampled coil-combined images 𝐴∗ 𝑥𝑢 . The optimization problem in Eq. 12
is solved by alternating projections for calibration-consistency (CC), data-consistency (DC) and
neural-network (CNN) consistency. Subnetworks are cascaded in series with data consistency
and calibration consistency projections. The data-consistency projection can be compactly
expressed as:
𝑓𝐶𝐶 {𝑥𝑢 } = 𝐺𝐹𝑥𝑢 (13)

In this multi-coil implementation, the input 𝑥𝑖𝑝 to the 𝑝𝑡ℎ subnetwork was formed as:

𝑓𝐷𝐶 {𝑓𝐶𝐶 {𝑥𝑢𝑛 }}, 𝑖𝑓 𝑝 = 1

𝑥𝑖𝑝 = { (14)
𝑓𝐷𝐶 {𝑓𝐶𝐶 {𝑓𝐷𝐶 {𝐶𝑝−1 (𝑓𝐷𝐶 {𝑓𝐶𝐶 {(… 𝑓𝐷𝐶 {𝐶1 (𝑓𝐷𝐶 {𝑓𝐶𝐶 {𝑥𝑢𝑛 }}; 𝜃1∗ )}; 𝜃𝑝−1
∗
)}, 𝑖𝑓 𝑝 > 1

Note that both calibration-consistency and neural-network consistency are followed by a data-
consistency layer.

Datasets

Public datasets. For demonstrations on single-coil data, three distinct datasets were used:
natural, T1-weighted brain, and T2-weighted brain images. The details are listed below.
1) Natural images: We assembled 5000 natural images from the validation set used during the
ImageNet Large Scale Visual Recognition Challenge 2011 (ILSVRC2011) (52). 4000 images
were used for training and 1000 images were used for validation. All images were either
cropped or zero-padded to yield consistent dimensions of 256x256. Color RGB images were
first converted to LAB color space, and the L-channel was extracted to obtain grayscale images.

2) T1-weighted images: We assembled a total of 6160 T1-weighted images (52 subjects) from
the MIDAS database (53). These images were divided into 4240 training images (36 subjects),
720 fine-tuning images (6 subjects) and 1200 testing images (10 subjects). In the training phase,
4000 images (34 subjects) were used for training while 240 images (2 subjects) were reserved
for validation. In the fine-tuning phase, 480 images (4 subjects) were used for fine-tuning and
240 images (2 subjects) were reserved for validation. There was no overlap between subjects
included in the training, validation and testing sets. T1-weighted images analyzed here were
collected on a 3T scanner via the following parameters: a 3D gradient-echo sequence,
TR=14ms, TE=7.7ms, flip angle=250, matrix size=256x176, 1 mm isotropic resolution.

3) T2-weighted images: We assembled a total of 5800 T2-weighted images (58 subjects) from
the MIDAS database (53). These images were divided into 4200 training images (42 subjects),
600 fine-tuning images (6 subjects) and 1000 testing images (10 subjects), with no subject
overlap between training, validation and testing sets. In the training phase, 4000 images (40
subjects) were used for training and 200 images (2 subjects) were used for validation. In the
fine-tuning phase, 400 images (4 subjects) were used for fine-tuning and 200 images (2
subjects) were used for validation. T2-weighted images analyzed here were collected on a 3T
scanner via the following parameters: a 2D spin-echo sequence, TR=7730ms, TE=80ms, flip
angle=900, matrix size=256x192, 1 mm isotropic resolution.

Multi-coil MR images. T1-weighted brain images from 10 subjects were acquired. Within each
subject, 60 central cross-sections containing sizeable amount of brain tissue were selected.
Images were then divided into 360 training images (6 subjects), 60 validation images (1 subject)
and 180 testing images (3 subjects), with no subject overlap. Images were collected on a 3T
Siemens Magnetom scanner (maximum gradient strength of 45mT/m and slew rate of 200
T/m/s) using a 32-channel receive-only head coil. The protocol parameters were: a 3D MP-
RAGE sequence, TR=2000ms, TE=5.53ms, flip angle=200, matrix size=256x192x80, 1 mm
isotropic resolution. Imaging protocol was approved by the local ethics committee at Bilkent
University and all participants provided written informed consent. To reduce computational
complexity, geometric-decomposition coil compression (GCC) was performed to reduce
number of coils from 32 to 8 (54).

Multi-coil natural images. To perform domain transfer from natural images to multi-coil MRI,
complex natural images were simulated from 2000 magnitude images in ImageNet by adding
sinusoidal phase at random spatial frequencies along each axis varying from – to +. The
amplitude of the sinusoids was normalized between 0 and 1. Fully-sampled multi-coil T1-
weighted acquisitions from 2 training subjects were selected to extract coil-sensitivity maps
using ESPIRiT (55). Each multi-coil complex natural image was then simulated by utilizing
coil-sensitivity maps of a randomly selected cross-section from the 2 reserved subjects (see
Supp. Figure 1 for sample multi-coil complex natural images).

Undersampling patterns. Images in each dataset were undersampled via variable-density

Poisson-disc sampling (50). All datasets were undersampled for varying acceleration factors
(R= 2, 4, 6, 8, 10). Fully sampled images were first Fourier transformed and then retrospectively
undersampled. To ensure reliability against mask selection, 100 unique undersampling masks
were generated and used during the training phase. A different set of 100 undersampling masks
were used during the testing phase.

Network training

We adopted a cascade of neural networks as inspired by (11). Five subnetworks were cascaded
in series. Each subnetwork contained an input layer, four convolutional layers and an output
layer. The input layer consisted of two channels for real imaginary parts of undersampled
images. Each convolution operation in the convolutional layers was passed through a rectified
linear unit (ReLU) activation. The hidden layers consisted of 64 channels. For single-coil data,
the output layer consisted of only a single channel for a magnitude reconstruction. For multi-
coil data, separate subnetworks were trained with real and imaginary parts of the reconstruction.

Subnetwork training. Subnetworks were trained via the back-propagation algorithm (56). In the
forward passes, a batch of 50 samples were passed through the network to calculate the
respective loss function. In the backward passes, network parameters were updated according
to the gradients of this function with respect to the parameters. The gradient of the loss function
with respect to parameters of the 𝑚𝑡ℎ hidden layer (𝜃𝑚 ) can be calculated using chain rule:

𝜕𝐿 𝜕𝐿 𝜕𝑜𝑙 𝜕𝑎𝑙 𝜕𝑜𝑙−1 𝜕𝑜𝑚 𝜕𝑎𝑚

= .... (15)
𝜕𝜃𝑚 𝜕𝑜𝑙 𝜕𝑎𝑙 𝜕𝜃𝑙 𝜕𝑎𝑙−1 𝜕𝑎𝑚 𝜕𝜃𝑚

where 𝑙 is the output layer of the network, 𝑎𝑙 is the output of the 𝑙 𝑡ℎ layer, and 𝑜𝑙 is the output
of the 𝑙 𝑡ℎ layer passed through the activation function. The parameters of the 𝑚𝑡ℎ layer were
only updated if the loss-function gradient flows through all subsequent layers (is non-zero).
Each subnetwork was trained individually for 20 epochs. In the training phase, the network
parameters were optimized using the ADAM optimizer with a learning rate of η=10 -4, decay
rate for first moment of gradient estimates of β1= 0.9 and decay rate for the second moment of
gradient estimate of β2=0.999 (57). Connection weights were L2-regularized with a
regularization parameter of 𝛾𝛷 =10-6.

Fine tuning. Networks formed by sequential training of the subnetworks were fine tuned. Here,
end-to-end fine tuning was performed on the entire neural-network architecture. To do this, the
gradients must be calculated through CNN, DC and CC blocks. The gradient flow through the
convolutional subnetworks that contain basic arithmetic operations and ReLU activation
functions are well known. The gradient flow through DC in Eq. 7 with respect to its input
𝐶(𝐴∗ 𝑥𝑢 ; 𝜃 ∗ ) is given as:

𝜕𝑓𝐷𝐶
= 𝐹−1 𝛬𝐹 (16)
𝜕𝐶(𝐴∗ 𝑥𝑢 ; 𝜃∗ )

due to linearity of the Fourier operator (𝐹 ). Similarly, the gradient flow through CC in Eq. 12
with respect to output of the preceding subnetwork 𝐶(𝐴∗ 𝑥𝑢 ; 𝜃 ∗ ) is given as:

𝜕𝑓𝐶𝐶 𝜕 𝑓𝐶𝐶 𝜕𝑓𝐷𝐶

= = 𝐺𝐹𝐹−1 𝛬𝐹 (17)
𝜕𝐶(𝐴∗ 𝑥𝑢 ; 𝜃∗ ) 𝜕𝑓𝐷𝐶 𝜕𝐶(𝐴∗ 𝑥𝑢 ; 𝜃∗ )
During the fine-tuning phase, the ADAM optimizer was used with identical parameters to those
used in subnetwork training, apart from a lower learning rate of 10-5 and a total of 100 epochs.
Performance analyses

Single-coil data. We first evaluated the performance of networks under implicit domain transfer
(i.e., without fine tuning). We reasoned that a network trained and tested in the same domain
should outperform networks trained and tested on different domains. To investigate this issue,
we reconstructed undersampled T1-weighted acquisitions using the ImageNet-trained and T2-
trained networks for varying acceleration factors (R=2, 4, 6, 8, 10). The reconstructions
obtained via these two networks were compared with reference reconstructions obtained from
the network trained directly on T1-weighted images. To ensure that our results were not biased
by the selection of a specific MR contrast as the test set, we also reconstructed undersampled
T2-weighted acquisitions using the ImageNet-trained and T1-trained networks. The
reconstructions obtained via these two networks were compared with reference reconstructions
obtained from the network trained directly on T2-weighted images.

Next, we evaluated the performance of network under explicit domain transfer (i.e., with fine
tuning). Networks were fine-tuned end-to-end in the testing domain. When T1-weighted images
were the testing domain, ImageNet-trained and T2-trained networks were fine-tuned using a
small set of T1-weighted images with size ranging in [0 100]. When T2-weighted images were
the testing domain, ImageNet-trained and T1-trained networks were fine-tuned using a small
set of T2-weighted images with size ranging in [0 100]. In both cases, the performance of fine-
tuned networks was compared with the networks trained and further fine-tuned end-to-end
directly in the testing domain on 100 images.

Reconstruction performance of a fine-tuned network likely depends on the number of both

training and fine-tuning images. To examine potential interaction between the number of
training and fine-tuning samples, separate networks were trained using training sets of varying
size in [500 4000]. Each network was then fine-tuned using sets of varying size in [0 100].
Performance was evaluated to determine the number of fine-tuning samples that are required to
achieve near-optimal performance for each separate size of training set. Optimal performance
was taken as the peak signal-to-noise ratio (PSNR) of a network trained directly in the testing
domain.

NN-based reconstructions were also compared to those obtained by conventional CS (2).

Single-coil CS reconstructions were implemented via a nonlinear conjugate gradient method.
Daubechies-4 wavelets were selected as the sparsifying transform. Parameter selection was
performed to maximize PSNR on the validation images from the fine-tuning set. Consequently,
an L1-regularization parameter of 10-3, and 80 iterations for T1-weighted acquisitions and 120
iterations for -weighted acquisitions were observed to yield near-optimal performance broadly
across R.

Multi-coil data. We also demonstrated the proposed approach on multi-coil MR images. For
this purpose, a network was trained on 2000 synthetic multi-coil complex natural images (see
Methods for details). The network was then fine-tuned using a set of multi-coil T1-weighted
images with varying size in [0 100]. This set was randomly selected from the training subjects.
Reconstruction performance was compared with networks trained using 360 T1-weighted multi-
coil MR images (6 subjects) and L1-SPIRiT (50). A POCS implementation of SPIRiT was used.
For each R, parameter selection was performed to maximize PSNR on validation images drawn
from the multi-coil MR image dataset. An interpolation kernel width of 7, a Tikhonov
regularization parameter of 10-2 for calibration, an L1-regularization parameter of 10-3 were
observed to yield near-optimal performance across R. Meanwhile, the optimal number of
iterations varied based on acceleration factor. For R= [2, 4, 6, 8, 10], the following number of
iterations= [20, 30, 45, 65, 80] were selected. The interpolation kernels optimized for SPIRiT
were used in the calibration-consistency layers of the networks that contained 5 consecutive CC
projections.

To quantitatively compared alternative methods, we measured the structural similarity index

(SSIM) and peak signal-to-noise ratio (PSNR) between the reconstructed and fully-sampled
reference images. For multi-coil data, the reference image was taken as the coil-combined
image obtained via weighted linear combination using coil sensitivity maps from ESPIRiT. The
training and testing of NN architectures were performed in the TensorFlow framework (58)
using 2 NVIDIA Titan X Pascal GPUs (12 GB VRAM). Single-coil CS reconstructions were
performed via libraries in the SparseMRI V0.2 toolbox available at
https://people.eecs.berkeley.edu/~mlustig/Software.html. Multi-coil CS reconstructions were
performed via libraries in the SPIRiT V0.3 toolbox available at
https://people.eecs.berkeley.edu/~mlustig/Software.html.
 Results

A network trained on the same type of images that it will later be tested on should outperform
networks trained and tested on different types of images. However, this performance difference
should diminish following successful domain transfer between the training and testing domains.
To test this prediction, we first investigated generalization performance for implicit domain
transfer (i.e., without fine tuning) in a single-coil setting. The training domain contained natural
images from the ImageNet database or T2-weighted images, and the testing domain contained
T1-weighted images. Figure 2 displays reconstructions of an undersampled T1-weighted
acquisition via the ImageNet-trained, T2-trained and T1-trained networks for R=4. As expected,
the T1-trained network yields sharper and more accurate reconstructions compared to the raw
ImageNet–trained and T2-trained networks. Next, we examined explicit domain transfer where
ImageNet-trained and T2-trained networks were fine-tuned. In this case, all networks yielded
visually similar reconstructions. Furthermore, when compared against conventional
compressive sensing (CS), all network models yielded superior performance. Figure 3 displays
reconstructions of an undersampled T1-weighted acquisition via the ImageNet-trained, T2-
trained and T1-trained networks, and CS for R=4. The ImageNet-trained network produces
images of similar visual quality to other networks and it outperforms CS in terms of image
sharpness and residual aliasing artifacts.

To corroborate these visual observations, reconstruction performance was quantitively assessed

for both implicit and explicit domain transfer across R=2-10. PSNR and SSIM measurements
across the test set are listed in Table 1 and Supp. Table 1. For implicit domain transfer, the T1-
trained networks outperform domain-transferred networks and CS consistently across all R. For
explicit domain transfer, the differences between the T1-trained and domain-transferred
networks diminish. Following fine-tuning, the average differences in (PSNR, SSIM) across R
between ImageNet and T1-trained networks diminish from (1.97dB, 3.80%) to (0.18dB,
0.20%), and difference between T2-trained and T1-trained networks diminish from (1.34dB,
2.20%) to (0.04dB, 0%). Furthermore, the domain-transferred networks outperform CS
consistently across R, by an average of 4.00dB PSNR and 9.9% SSIM.

Next, we repeated the analyses for implicit and explicit domain transfer when the testing
domain contained T2-weighted images. Supp. Figure 2 displays reconstructions of an
undersampled T2-weighted acquisition via the ImageNet-, T1- and T2-trained networks for
acceleration factor R=4. Again, the network trained directly in the testing domain (T2-weighted)
outperforms domain transferred networks. After fine tuning with as few as 20 images, the
domain-transferred networks yield visually similar reconstructions to the T2-trained network.

PSNR and SSIM measurements on T2-weighted reconstructions across the test set are listed in
Supp. Table 2. Following fine-tuning, average (PSNR, SSIM) differences between ImageNet
and T2-trained networks diminish from (1.23dB, 3.40%) to (0.19dB, 0.40%), and difference
between T1-trained and T2-trained networks diminish from (1.14dB, 2.80%) to (0.14dB, -
0.20%). Across R, the domain-transferred networks also outperform CS by 5.21dB PSNR and
12.5% SSIM.

Reconstruction performance of domain-transferred networks may depend on the sizes of both

training and fine-tuning sets. To examine interactions between the number of training (Ntrain)
and fine-tuning (Ntune) samples, we trained networks using training sets in the range [500 4000]
and fine-tuning sets in the range [0 100]. Figure 4 shows average PSNR values for a reference
T1-trained network trained on 4000 and fine-tuned on 100 images, and domain transferred
networks for R=2-10. Without fine-tuning, the T1-trained network outperforms both domain-
transferred networks. As the number of fine-tuning samples increases, the PSNR differences
decay gradually to a negligible level. Consistently across R, domain-transferred networks
trained on smaller training sets require more fine-tuning samples to yield similar performance.

Figure 5 displays the number of fine-tuning samples required for the PSNR values for
ImageNet-trained networks to converge for R=2-10. Convergence was taken as the number of
fine-tuning samples where the percentage change in PSNR by incrementing number of fine-
tuning samples fell below 0.05% of PSNR for the T1-trained network. Consistently across R,
networks trained on fewer samples require more fine-tuning samples for convergence.
However, the required number of fine-tuning samples is greater for higher R. Averaged across
R, Ntune=72 for Ntrain=500, Ntune=57 for Ntrain=1000, Ntune=58 for Ntrain=2000, Ntune=44 for
Ntrain=4000.

We also examined interactions between the number of training and fine-tuning samples when
the target domain contained T2-weighted images. Supp. Figure 4 shows average PSNR values
for a reference T2-trained network trained on 4000 and fine-tuned on 100 images, and domain
transferred networks for R=2-10. Similar to the case of T1-weighted images, domain-transferred
networks trained on smaller sets require more fine-tuning samples to yield comparable
performance. Supp. Figure 5 displays the number of fine-tuning samples required for
convergenge of ImageNet-trained networks. Averaged across R=2-10, Ntune=66 for Ntrain=500,
Ntune=46 for Ntrain=1000, Ntune=51 for Ntrain=2000, Ntune=43 for Ntrain=4000.

Next, we demonstrated the proposed approach on multi-coil T1-weighted images. We compared

ImageNet- and T1-trained networks at R=2-10. Figure 6 displays average PSNR values for the
T1-trained network (trained and fine-tuned on 360 images) and ImageNet-trained network
(trained on 2000 multi-coil natural images and fine-tuned on [0,100] T1-weighted images). As
Ntune increases, the PSNR differences between T1- and ImageNet-trained networks start
diminishing. Figure 7 displays the number of fine-tuning samples required for the PSNR values
for ImageNet-trained networks to converge. Averaged across R=2-10, ImageNet-trained
networks require Ntune=31 for convergence. We also compared the proposed transfer learning
approach with L1-regularized SPIRiT. Figure 8 shows representative reconstructions obtained
via the ImageNet-trained network, T1-trained network and SPIRiT for R=10. The ImageNet-
trained network produces images of similar visual quality to the T1-trained network, and it
outperforms SPIRiT in terms of residual aliasing artifacts.

Quantitative assessment of multi-coil reconstructions for the ImageNet-trained network, T1-

trained network and SPIRiT across R=2-10 are listed in Table 2. For implicit domain transfer,
the T1-trained network performs better than the ImageNet-trained network. Following fine-
tuning, the average differences in (PSNR, SSIM) across R between ImageNet and T1-trained
networks diminish from (1.92dB, 1.00%) to (0.56dB, 0.10%). Furthermore, the ImageNet-
trained network outperforms SPIRiT in all cases, except for R=2 where SPIRiT yields higher
PSNR and SSIM, and R=4 where the two methods yield similar PSNR. On average across R,
the ImageNet-trained network improves performance over SPIRiT by 0.67dB PSNR and 0.50%
SSIM.
 Discussion

Neural networks for MRI reconstruction involve many free parameters to be learnt, so an
extensive amount of training samples is typically needed (59). In theory, network performance
should be optimized by drawing the training and testing samples from the same domain,
acquired under a common MRI protocol. In practice, however, compiling large public datasets
can require coordinated efforts among multiple imaging centers, and so such datasets are rare.
As an alternative, several recent studies trained neural networks on a collection of multi-contrast
images (19). When needed, data augmentation procedures were used to further expand the
training dataset (11,12). While these approaches gather more samples for training, it remains
unclear how well a network trained on images acquired with a specific type of tissue contrast
generalizes to images acquired with different contrasts. Thus, variability in MR contrasts can
lead to suboptimal reconstruction performance.

Here, we first questioned the generalizability of neural network models across different
contrasts. We find that a network trained on MR images of a given contrast (e.g. T1-weighted)
yields suboptimal reconstructions on images of a different contrast (e.g. T2-weighted). This
confirms that the best strategy is to train and test networks in the same domain. Yet, it may not
be always feasible to gather a large collection of images from a desired contrast. To address the
problem of data scarcity, we proposed a transfer-learning approach for accelerated MRI. The
proposed approach trains neural networks using training samples from a large public dataset of
natural images. The network is then fine-tuned end-to-end using only few tens of MR images.
Reconstructions obtained via the ImageNet-trained network are of nearly identical quality to
reconstructions obtained by networks trained directly in the testing domain using thousands of
MR images.

Several recent studies have considered domain transfer to enhance performance in NN-based
MRI reconstruction (18,29,30,42). A group of studies have aimed to perform implicit domain
transfer across MRI contrasts without fine-tuning. One proposed method was to train networks
on MR images in a given contrast, and then to directly use the trained networks on images of
different contrasts (18). While this method yields successful reconstructions, our results suggest
that network performance can be further boosted with additional fine-tuning in the testing
domain. Another method to enhance generalizability was to compound datasets containing a
mixture of distinct MRI contrasts during network training (29). This approach enforces the
network to better adapt to variations in tissue contrast. Yet, in the absence of contrast-specific
fine-tuning, networks may deliver suboptimal performance for some individual contrasts.

A second group of studies have attempted explicit domain transfer across training and testing
domains via fine-tuning. A recent proposed method trained a deep residual network to remove
streaking artifacts from CT images, and the trained network was then used to suppress aliasing
artifacts in projection-reconstruction MRI (30). This method leverages the notion that the
characteristic structure of artifacts due to polar sampling should be similar in CT and MRI.
Here, we considered random sampling patterns on a Cartesian grid, and therefore, the domain
transfer method proposed in (30) is not directly applicable to our reconstructions that possess
incoherent artifacts. Another recent, independent effort examined the reliability of
reconstructions from a variational network to deviations in undersampling patterns and SNR
between the training and testing domains (42). Mismatch in patterns or SNR between the two
domains caused suboptimal performance even for modest acceleration factors. They also
assessed the generalization capability by performing implicit domain transfer between PD-
weighted knee images with and without fat suppression. A network trained on PD-weighted
knee images without fat suppression was observed to yield relatively poor reconstructions of
images with fat suppression and vice versa. Consistent with these observations, we also find
that, without fine-tuning, networks trained on MR images of a given contrast (e.g. T1-weighted)
do not generalize well to images of a different contrast (e.g. T2-weighted). That said, a distinct
contribution of our work was to address the issue of data scarcity by training a network in a
domain with ample data, and transferring the network to a domain with fewer samples.

An alternative approach proposed to train neural networks for MRI reconstruction with small
datasets is Robust artificial‐neural‐networks for k‐space interpolation, RAKI (13). This
previous method aims to train a neural network for each individual subject that learns to
synthesize missing k-space samples from acquired data. Unlike traditional k-space parallel
imaging methods (50,60), a nonlinear interpolation kernel was estimated from central
calibration data. Such nonlinear interpolation was shown to boost reconstruction performance
beyond linear methods. However, RAKI might yield suboptimal performance when the optimal
interpolation kernel shows considerable variation across k-space. Our proposed architecture for
multi-coil reconstructions leverages a linear interpolation kernel, so the output of calibration-
consistency blocks in our network can manifest similar reconstruction errors. Yet, the remaining
CNN blocks are trained to recover fully-sampled reference images given images with residual
artifacts at the output of CC blocks.

Here, we demonstrated domain transfer based on a cascade architecture with multiple CNNs
interleaved with data- and calibration-consistency layers. The proposed approach might
facilitate the use of neural networks for MRI reconstruction in applications where data are
relatively scarce. It might also benefit other types of architectures that have been proposed for
accelerated MRI (16,19,41), in particular architectures that require extensive datasets for
adequate training (12,45). Here, the calibration-consistency projections were based on the
SPIRiT method. These projections can also be replaced with other k-space methods for parallel
imaging such as GRAPPA or RAKI. Note that the current study examined the generalization
capability of networks trained on natural images to T1-weighted and T2-weighted images of the
brain. ImageNet-trained networks could also be beneficial for reconstruction of MR images
acquired with more specialized contrasts such as angiograms, and images acquired in other
body parts.
 Tables

Table 1. Reconstruction quality for single-coil T1-weighted images undersampled at R= 2, 4,

6, 8, 10. Reconstructions were performed via ImageNet-trained, T1-trained and T2-trained
networks. PSNR and SSIM values are reported as mean±standard deviation across test images.
Results are shown for raw networks trained on 2000 training images (raw), and fine-tuned
networks tuned with few tens of T1-weighted images (tuned).

ImageNet-trained T1-trained T2-trained

R=2

PSNR SSIM PSNR SSIM PSNR SSIM

Raw 40.33 ± 3.42 0.96 ± 0.02 40.65 ± 3.07 0.97 ± 0.01 40.15 ± 3.14 0.96 ± 0.01
Tuned 42.81 ± 3.32 0.97 ± 0.01 42.37 ± 3.25 0.97 ± 0.01 42.75 ± 3.22 0.97 ± 0.01
ImageNet-trained T1-trained T2-trained
PSNR SSIM PSNR SSIM PSNR SSIM
R=4

Raw 34.07 ±3.19 0.89 ± 0.03 34.87 ± 2.90 0.91 ± 0.02 33.26 ± 3.23 0.90 ± 0.03
Tuned 35.85 ± 3.03 0.93± 0.03 36.09 ± 3.19 0.93 ± 0.03 35.95 ±3.03 0.93 ± 0.03
ImageNet-trained T1-trained T2-trained
R=6

PSNR SSIM PSNR SSIM PSNR SSIM

Raw 29.42 ± 3.59 0.84 ± 0.04 32.34 ± 2.95 0.89 ± 0.03 30.48 ± 3.22 0.86 ± 0.03
Tuned 33.47 ± 3.11 0.90± 0.03 33.90 ± 3.26 0.90 ± 0.04 33.63 ± 3.09 0.90 ± 0.03
ImageNet-trained T1-trained T2-trained
R=8

PSNR SSIM PSNR SSIM PSNR SSIM

Raw 27.28 ± 3.77 0.81 ± 0.04 30.07 ± 3.18 0.86 ± 0.03 28.42 ± 3.14 0.83 ± 0.04
Tuned 32.14 ± 3.22 0.89 ± 0.04 32.21 ± 3.32 0.89 ± 0.04 32.17 ± 3.45 0.89 ± 0.04
ImageNet-trained T1-trained T2-trained
R=10

PSNR SSIM PSNR SSIM PSNR SSIM

Raw 25.82 ± 3.85 0.79 ± 0.05 28.84 ± 3.43 0.85 ± 0.04 27.72 ± 3.30 0.82 ± 0.04
Tuned 30.93 ± 3.40 0.87 ± 0.04 31.53 ± 3.38 0.88 ± 0.04 31.42 ± 3.28 0.88 ± 0.04
Table 2. Reconstruction quality for multi-coil T1-weighted images undersampled at R= 2, 4, 6,
8, 10. Reconstructions were performed via ImageNet-trained and T1-trained networks as well
as SPIRiT. PSNR and SSIM values are reported as mean±standard deviation across test images.
Results are shown for raw networks trained on 2000 training images (raw), and fine-tuned
networks tuned with few tens of T1-weighted images (tuned).

ImageNet-trained T1-trained SPIRiT

R=2

PSNR SSIM PSNR SSIM PSNR SSIM

Raw 49.48±1.73 0.995±.001
50.46±1.67 0.996±.002 50.47±1.68 0.996±.002
Tuned 50.09±1.64 0.996±.002
ImageNet-trained T1-trained SPIRiT
R=4

PSNR SSIM PSNR SSIM PSNR SSIM

Raw 43.54±1.74 0.985±.006
45.36±1.75 0.989±.004 44.60±1.75 0.987±.004
Tuned 44.76±1.76 0.989±.004
ImageNet-trained T1-trained SPIRiT
R=6

PSNR SSIM PSNR SSIM PSNR SSIM

Raw 39.55±1.90 0.970±.010
42.06±1.85 0.981±.006 40.62±1.73 0.975±.007
Tuned 41.40±1.95 0.980±.007
ImageNet-trained T1-trained SPIRiT
R=8

PSNR SSIM PSNR SSIM PSNR SSIM

Raw 36.78±1.72 0.954±.014
39.14±1.75 0.971±.009 37.11±1.71 0.961±.012
Tuned 38.45±1.77 0.969±.010
ImageNet-trained T1-trained SPIRiT
R=10

PSNR SSIM PSNR SSIM PSNR SSIM

Raw 34.27±1.72 0.943±.016
36.19±1.85 0.960±.012 34.23±1.72 0.948±.017
Tuned 35.68±1.89 0.958±.014
 Figures

Figure 1. Proposed transfer-learning approach for NN-based reconstructions of multi-coil (Nc

coils) undersampled acquisitions. A deep architecture with multiple subnetworks is used. The
subnetworks consist of calibration consistency “CC” and CNN “Conv” blocks, each followed
by a data consistency block “DC”. (a) Each subnetwork is trained sequentially to reconstruct
synthetic multi-coil natural images from ImageNet, given zero-filled Fourier reconstructions of
their undersampled versions. Due to differences in the characteristics of natural and MR images,
the ImageNet-trained network will yield suboptimal performance when directly tested on MR
images. (b) For domain transfer, the ImageNet-trained network is fine-tuned end-to-end in the
testing domain using few tens of images. This approach enables successful domain transfer
between natural and MR images.
Figure 2. Representative reconstructions of a T1-weighted acquisition at acceleration factor
R=4. Reconstructions were performed via the Zero-filled Fourier method (ZF), and ImageNet-
trained, T2-trained, and T1-trained networks. (a) Reconstructed images and error maps for raw
networks (see colorbar). (b) Reconstructed images and error maps for fine-tuned networks. The
fully-sampled reference image is also shown. Network training was performed on a training
dataset of 2000 images and fine-tuned on a sample of 20 T1-weighted images. Following fine-
tuning with few tens of samples, ImageNet-trained and T2-trained networks yield
reconstructions of highly similar quality to the T1-trained network.
Figure 3. Reconstructions of a T1-weighted acquisition with R=4 via ZF, conventional
compressed-sensing (CS), and ImageNet-trained, T1-trained and T2-trained networks along
with the fully-sampled reference image. Error maps for each reconstruction are shown below
(see colorbar). Networks were trained on 2000 images and fine-tuned on 20 images acquired
with the test contrast. The domain-transferred networks maintain nearly identical performance
to the networks trained directly in the testing domain. Furthermore, the domain transferred-
networks reconstructions outperform conventional CS in terms of image sharpness and residual
aliasing artifacts.
Figure 4. Reconstruction performance was evaluated for undersampled T1-weighted
acquisitions. Average PSNR values across T1-weighted validation images were measured for
the T1-trained network (trained on 4k images and fine-tuned on 100 images), ImageNet-trained
networks (trained on 500, 1000, 2000, or 4000 images), and T2-trained network (trained on
4000 images). Results are plotted as a function of number of fine-tuning samples for
acceleration factors (a) R=2, (b) R= 4, (c) R = 6, (d) R = 8, and (e) R= 10. Without fine-tuning,
the T1-trained network outperforms all domain-transferred networks. As the number of fine-
tuning samples increases, the PSNR differences decay gradually to a negligible level. Domain-
transferred networks trained on fewer samples require more fine-tuning samples to yield similar
performance consistently across R.
Figure 5. Number of fine-tuning samples required for the PSNR values for ImageNet-trained
networks to converge. Average PSNR values across T1-weighted validation images were
measured for the ImageNet-trained networks trained on (a) 500, (b) 1000, (c) 2000, and (d)
4000 images. Convergence was taken as the number of fine-tuning samples where the
percentage change in PSNR by incrementing N tune fell below 0.05% of the average PSNR for
the T1-trained network (see Figure 4). Domain-transferred networks trained on fewer samples
require more fine-tuning samples for the PSNR values to converge. Furthermore, at higher
values of R, more fine-tuning samples are required for convergence.
Figure 6. Reconstruction performance was evaluated for undersampled multi-coil T1-weighted
acquisitions. Average PSNR values across T1-weighted validation images were measured for
the T1-trained network (trained and fine-tuned on 360 images), and ImageNet-trained network
trained on 2000 images. Results are plotted as a function of number of fine-tuning samples for
acceleration factors (a) R=2, (b) R= 4, (c) R = 6, (d) R = 8, and (e) R= 10. Without fine-tuning,
the T1-trained network outperforms the domain-transferred network. As the number of fine-
tuning samples increases, the PSNR differences decay gradually to a negligible level.
Figure 7. Number of fine-tuning samples required for the PSNR values for ImageNet-trained
networks to converge. Average PSNR values across T1-weighted validation images were
measured for the ImageNet-trained network trained on 2000 images. Convergence was taken
as the number of fine-tuning samples where the percentage change in PSNR by incrementing
Ntune fell below 0.05% of the average PSNR for the T1-trained network (see Figure 6). At
higher values of R, more fine-tuning samples are required for convergence.
Figure 8. Representative reconstructions of a multi-coil T1-weighted acquisition at acceleration
factor R=10. Reconstructions were performed via ZF, ImageNet-trained and T1-trained
networks, and SPIRiT (top row). Corresponding error maps are also shown (see colorbar;
bottom row) along with the fully-sampled reference (top row). Network training was performed
on a training dataset of 2000 images and fine-tuned on a sample of 20 T1-weighted images. The
ImageNet-trained network maintains similar performance to the T1-trained network trained
directly on the images from the test domain. Furthermore, the domain-transferred network
outperforms conventional SPIRiT in terms of residual aliasing artifacts.
 References

1. Block KT, Uecker M, Frahm J. Undersampled radial MRI with multiple coils. Iterative image
reconstruction using a total variation constraint. Magn. Reson. Med. 2007;57:1086–1098 doi:
10.1002/mrm.21236.
2. Lustig M, Donoho D, Pauly JM. Sparse MRI: The application of compressed sensing for
rapid MR imaging. Magn. Reson. Med. 2007;58:1182–1195 doi: 10.1002/mrm.21391.
3. Çukur T, Lustig M, Nishimura DG. Improving non-contrast-enhanced steady-state free
precession angiography with compressed sensing. Magn. Reson. Med. 2009;61:1122–1131
doi: 10.1002/mrm.21907.
4. Jung H, Ye JC. Performance evaluation of accelerated functional MRI acquisition using
compressed sensing. In: IEEE International Symposium on Biomedical Imaging: From Nano
to Macro. ; 2009. pp. 702–705. doi: 10.1109/isbi.2009.5193144.
5. Menzel MI, Tan ET, Khare K, et al. Accelerated diffusion spectrum imaging in the human
brain using compressed sensing. Magn. Reson. Med. 2011;66:1226–1233 doi:
10.1002/mrm.23064.
6. Doneva M, Börnert P, Eggers H, Stehning C, Sénégas J, Mertins A. Compressed sensing
reconstruction for magnetic resonance parameter mapping. Magn. Reson. Med.
2010;64:1114–1120 doi: 10.1002/mrm.22483.
7. Ravishankar S, Bresler Y. MR image reconstruction from highly undersampled k-space data
by dictionary learning. IEEE Trans. Med. Imaging 2011;30:1028–1041 doi:
10.1109/TMI.2010.2090538.
8. Khare K, Hardy CJ, King KF, Turski PA, Marinelli L. Accelerated MR imaging using
compressive sensing with no free parameters. Magn. Reson. Med. 2012;68:1450–1457 doi:
10.1002/mrm.24143.
9. Weller DS, Ramani S, Nielsen J-F, Fessler JA. Monte Carlo SURE-based parameter
selection for parallel magnetic resonance imaging reconstruction. Magn. Reson. Med.
2014;71:1760–1770 doi: 10.1002/mrm.24840.
10. Jun Y, Eo T, Kim T, Jang J, Hwang D. Deep Convolutional Neural Network for Acceleration
of Magnetic Resonance Angiography (MRA). In: 25th Annual Meeting of ISMRM. ; 2017. p.
0686.
11. Schlemper J, Caballero J, Hajnal J V., Price A, Rueckert D. A Deep Cascade of
Convolutional Neural Networks for MR Image Reconstruction. In: International Conference on
Information Processing in Medical Imaging. ; 2017. pp. 647–658.
12. Yu S, Dong H, Yang G, et al. DAGAN: Deep de-aliasing generative adversarial networks
for fast compressed sensing MRI reconstruction. IEEE Trans. Med. Imaging 2018;37:1310–
1321.
13. Akçakaya M, Moeller S, Weingärtner S, Uğurbil K. Scan-specific robust artificial-neural-
networks for k-space interpolation (RAKI) reconstruction: Database-free deep learning for fast
imaging. Magn Reson Med 2019:81 doi: 10.1002/mrm.27420.
14. Lee D, Yoo J, Ye JC. Compressed sensing and Parallel MRI using deep residual learning.
In: 25th Annual Meeting of ISMRM. ; 2017. p. 0641.
15. Kwon K, Kim D, Park H. A parallel MR imaging method using multilayer perceptron. Med.
Phys. 2017;44:6209–6224 doi: 10.1002/mp.12600.
16. Yang Y, Sun J, Li H, Xu Z. Deep ADMM-Net for Compressive Sensing MRI. In: Advances
in Neural Information Processing Systems. ; 2016.
17. Zhu B, Liu JZ, Rosen BR, Rosen MS. Neural Network MR Image Reconstruction with
AUTOMAP: Automated Transform by Manifold Approximation. In: 25th Annual Meeting of
ISMRM. ; 2017. p. 0640. doi: 10.1137/S003614450343200X.
18. Zhu B, Liu JZ, Rosen BR, Rosen MS. Image reconstruction by domain transform manifold
learning. Nature 2018;555:487–492 doi: 10.1017/CCOL052182303X.002.
19. Hyun CM, Kim HP, Lee SM, Lee S, Seo JK. Deep learning for undersampled MRI
reconstruction. Phys. Med. Biol. 2018;63:135007.
20. Quan TM, Nguyen-Duc T, Jeong W-K. Compressed sensing MRI reconstruction with cyclic
loss in generative adversarial networks. IEEE Trans. Med. Imaging 2018;37:1488–1497.
21. Akçakaya M, Moeller S, Weingärtner S, Uğurbil K. Scan-specific robust artificial-neural-
networks for k-space interpolation (RAKI) reconstruction: Database-free deep learning for fast
imaging. Magn. Reson. Med. 2019;81:439–453 doi: 10.1002/mrm.27420.
22. Akçakaya M, Moeller S, Weingärtner S, Ugurbil K. Scan-specific Robust Artificial-neural-
networks for k-space Interpolation (RAKI): Database-free Deep Learning Reconstruction for
Fast Imaging. In: 26th Annual Meeting of ISMRM. ; 2018. p. 0576.
23. Pawar K, Chen Z, Shah N, Egan G. ReconNet: A Deep Learning Framework for
Transforming Image Reconstruction into Pixel Classification. In: 26th Annual Meeting of
ISMRM. ; 2018. p. 0574.
24. Fu Z, Mandava S, Keerthivasan M, Martin D, Altbach M, Bilgin A. A Multi-scale Deep
ResNet for Radial MR Parameter Mapping. In: 26th Annual Meeting of ISMRM. ; 2018. p. 0568.
25. Chang P, Liu M, Chow D, et al. Inception-CS: Deep Learning For Sparse MR
Reconstruction in Glioma Patients. In: 26th Annual Meeting of ISMRM. ; 2018. p. 0571.
26. Zhu B, Bilgic B, Liao C, Rosen B, Rosen M. Deep learning MR reconstruction with
Automated Transform by Manifold Approximation (AUTOMAP) in real-world acquisitions with
imperfect training. In: 26th Annual Meeting of ISMRM. ; 2018. p. 0572.
27. Lee J, Ye JC. Deep convolutional framelet neural network for reference-free EPI ghost
correction. In: 26th Annual Meeting of ISMRM. ; 2018. p. 0575.
28. Moeller S, Weingärtner S, Ugurbil K, Akcakaya M. Application of a Scan-Specific Deep
Learning Reconstruction to Multiband/SMS Imaging. In: 26th Annual Meeting of ISMRM. ;
2018. p. 0577.
29. Wang S, Su Z, Ying L, et al. Accelerating magnetic resonance imaging via deep learning.
In: IEEE 13th International Symposium on Biomedical Imaging (ISBI). ; 2016. pp. 514–517.
doi: 10.1109/ISBI.2016.7493320.
30. Han Y, Yoo J, Kim HH, Shin HJ, Sung K, Ye JC. Deep learning with domain adaptation for
accelerated projection-reconstruction MR. Magn. Reson. Med. 2018;80:1189–1205 doi:
10.1002/mrm.27106.
31. Han Y, Ye JC. k-Space Deep Learning for Accelerated MRI. arXiv Prepr. 2018.
32. Kwon K, Kim D, Seo H, Cho J, Kim B, Park H. Learning-based Reconstruction using
Artificial Neural Network for Higher Acceleration. In: 24th Annual Meeting of ISMRM. ; 2016.
p. 1801.
33. Gong E, Zaharchuk G, Pauly J. Improving the PI+CS Reconstruction for Highly
Undersampled Multi-contrast MRI using Local Deep Network. In: 25th Annual Meeting of
ISMRM. ; 2017. p. 5663.
34. Wang S, Huang N, Zhao T, Yang Y, Ying L, Liang D. 1D Partial Fourier Parallel MR imaging
with deep convolutional neural network. In: 25th Annual Meeting of ISMRM. ; 2017. p. 0642.
35. Chen F, Taviani V, Malkiel I, et al. Variable-Density Single-Shot Fast Spin-Echo MRI with
Deep Learning Reconstruction by Using Variational Networks. Radiology 2018;289:366–373
doi: 10.1148/radiol.2018180445.
36. Cheng JY, Chen F, Alley MT, Pauly JM, Vasanawala SS. Highly Scalable Image
Reconstruction using Deep Neural Networks with Bandpass Filtering. arXiv Prepr. 2018.
37. Mardani M, Gong E, Cheng J, et al. Deep Generative Adversarial Neural Networks for
Compressed Sensing (GANCS) Automates MRI. In: 26th Annual Meeting of ISMRM. ; 2018.
p. 0573.
38. Joseph Y. Cheng, Morteza Mardani, Marcus T. Alley, John M. Pauly and SSV.
DeepSPIRiT: Generalized Parallel Imaging using Deep Convolutional Neural Networks. In:
25th Annual Meeting of ISMRM. ; 2018. p. 0570.
39. Mardani M, Gong E, Cheng JY, et al. Deep Generative Adversarial Neural Networks for
Compressive Sensing (GANCS) MRI. IEEE Trans. Med. Imaging 2018:1–1 doi:
10.1109/TMI.2018.2858752.
40. Hammernik K, Knoll F, Sodickson DK, Pock T. Learning a Variational Model for
Compressed Sensing MRI Reconstruction. In: 24th Annual Meeting of ISMRM. ; 2016. p. 1088.
41. Hammernik K, Klatzer T, Kobler E, et al. Learning a Variational Network for Reconstruction
of Accelerated MRI Data. Magn. Reson. Med. 2017;79:3055–3071.
42. Knoll F, Hammernik K, Kobler E, Pock T, Recht MP, Sodickson DK. Assessment of the
generalization of learned image reconstruction and the potential for transfer learning. Magn.
Reson. Med. 2019;81:116–128 doi: 10.1002/mrm.27355.
43. Hammernik K, Klatzer T, Kobler E, et al. Learning a variational network for reconstruction
of accelerated MRI data. Magn. Reson. Med. 2018;79:3055–3071 doi: 10.1002/mrm.26977.
44. Krizhevsky A, Sutskever I, Hinton GE. ImageNet Classification with Deep Convolutional
Neural Networks. In: Pereira F, Burges CJC, Bottou L, Weinberger KQ, editors. Advances in
Neural Information Processing Systems 25. Curran Associates, Inc.; 2012. pp. 1097–1105.
45. Mardani M, Gong E, Cheng JY, et al. Deep generative adversarial neural networks for
compressive sensing MRI. IEEE Trans. Med. Imaging 2019;38:167–179.
46. Donahue J, Jia Y, Vinyals O, et al. DeCAF: A Deep Convolutional Activation Feature for
Generic Visual Recognition. In: Xing EP, Jebara T, editors. Proceedings of the 31st
International Conference on Machine Learning. Vol. 32. Proceedings of Machine Learning
Research. Bejing, China: PMLR; 2014. pp. 647–655.
47. Yosinski J, Clune J, Bengio Y, Lipson H. How transferable are features in deep neural
networks? In: Ghahramani Z, Welling M, Cortes C, Lawrence ND, Weinberger KQ, editors.
Advances in Neural Information Processing Systems 27. Curran Associates, Inc.; 2014. pp.
3320–3328.
48. Dar SUH, Çukur T. Transfer learning for reconstruction of accelerated MRI acquisitions via
neural networks. In: 26th Annual Meeting of ISMRM. ; 2018. p. 0569.
49. Shahdloo M, Ilicak E, Tofighi M, Saritas EU, Cetin AE, Cukur T. Projection onto Epigraph
Sets for Rapid Self-Tuning Compressed Sensing MRI. IEEE Trans. Med. Imaging 2018:1–1
doi: 10.1109/TMI.2018.2885599.
50. Lustig M, Pauly JM. SPIRiT: Iterative self-consistent parallel imaging reconstruction from
arbitrary k-space. Magn. Reson. Med. 2010;64:457–471 doi: 10.1002/mrm.22428.
51. Bauschke HH, Borwein JM. On Projection Algorithms for Solving Convex Feasibility
Problems. SIAM Rev. 1996;38:367–426 doi: 10.1137/S0036144593251710.
52. Russakovsky O, Deng J, Su H, et al. ImageNet Large Scale Visual Recognition Challenge.
Int. J. Comput. Vis. 2015;115:211–252 doi: 10.1007/s11263-015-0816-y.
53. Bullitt E, Zeng D, Gerig G, et al. Vessel tortuosity and brain tumor malignancy: A blinded
study. Acad. Radiol. 2005;12:1232–1240 doi: 10.1016/j.acra.2005.05.027.
54. Zhang T, Pauly JM, Vasanawala SS, Lustig M. Coil compression for accelerated imaging
with Cartesian sampling. Magn. Reson. Med. 2013;69:571–82 doi: 10.1002/mrm.24267.
55. Uecker M, Lai P, Murphy MJ, et al. ESPIRiT-an eigenvalue approach to autocalibrating
parallel MRI: Where SENSE meets GRAPPA. Magn. Reson. Med. 2014;71:990–1001 doi:
10.1002/mrm.24751.
56. Rumelhart DE, Hinton GE, Williams RJ. Learning representations by back-propagating
errors. Nature 1986;323:533–536 doi: 10.1038/323533a0.
57. Kingma DP, Ba JL. Adam: a Method for Stochastic Optimization. In: International
Conference on Learning Representations. ; 2015. doi:
http://doi.acm.org.ezproxy.lib.ucf.edu/10.1145/1830483.1830503.
58. Abadi M, Barham P, Chen J, et al. TensorFlow: Large-Scale Machine Learning on
Heterogenous Distributed Systems. In: 12th USENIX Symposium on Operating Systems
Design and Implementation (OSDI ’16). ; 2016. pp. 265–284.
59. Domingos P. A few useful things to know about machine learning. Commun. ACM
2012;55:78 doi: 10.1145/2347736.2347755.
60. Griswold MA, Jakob PM, Heidemann RM, et al. Generalized autocalibrating partially
parallel acquisitions (GRAPPA). Magn. Reson. Med. 2002;47:1202–1210 doi:
10.1002/mrm.10171.
 Supplementary Materials
Table of Contents
1. Supplementary Figure 1. Sample synthetic, complex, multi-coil natural images
2. Supplementary Figure 2. T2-trained vs domain transferred-networks for T2
reconstructions
3. Supplementary Figure 3. NN-based vs compressive sensing (CS) methods for T2
reconstructions
4. Supplementary Figure 4. Interaction between number of training and fine-tuning
samples for ImageNet-trained networks in T2 reconstructions
5. Supplementary Figure 5. Required number of fine-tuning samples for convergence in
T2 reconstructions via the ImageNet-trained network
6. Supplementary Table 1. Quality of T1 reconstructions
7. Supplementary Table 2. Quality of T2 reconstructions
Supplementary Figure 1. Representative synthetic complex multi-coil natural images.
Complex multi-coil natural images were simulated from magnitude images in ImageNet (see
Methods for details). Magnitude and phase of two sets of simulated images (a and b) are shown
along with their reference magnitude images.
Supplementary Figure 2. Representative reconstructions of a T2-weighted acquisition at
acceleration factor R=4. Reconstructions were performed via the Zero-filled Fourier method
(ZF), and ImageNet-trained, T2-trained, and T1-trained networks. (a) Reconstructed images and
error maps for raw networks (see colorbar). (b) Reconstructed images and error maps for fine-
tuned networks. The fully-sampled reference image is also shown. Network training was
performed on a training dataset of 2000 images and fine-tuned on a sample of 20 T2-weighted
images. Following fine-tuning with few tens of samples, ImageNet-trained and T1-trained
networks yield reconstructions of highly similar quality to the T1-trained network.
Supplementary Figure 3. Reconstructions of a T2-weighted acquisition with R=4 via ZF,
conventional compressed-sensing (CS), and ImageNet-trained, T1-trained and T2-trained
networks along with the fully-sampled reference image. Error maps for each reconstruction are
shown below (see colorbar). Networks were trained on 2000 images and fine-tuned on 20
images acquired with the test contrast. The domain-transferred networks maintain nearly
identical performance to the networks trained directly in the testing domain. Furthermore, the
domain transferred-networks reconstructions outperform conventional CS in terms of image
sharpness and residual aliasing artifacts.
Supplementary Figure 4. Reconstruction performance was evaluated for undersampled T2-
weighted acquisitions. Average PSNR values across T2-weighted validation images were
measured for the T1-trained network (trained on 4k images and fine-tuned on 100 images),
ImageNet-trained networks (trained on 500, 1000, 2000, or 4000 images), and T1-trained
network (trained on 4000 images). Results are plotted as a function of number of fine-tuning
samples for acceleration factors (a) R=2, (b) R= 4, (c) R = 6, (d) R = 8, and (e) R= 10. As the
number of fine-tuning samples increases, the PSNR differences decay gradually to a negligible
level. Domain-transferred networks trained on fewer samples require more fine-tuning samples
to yield similar performance consistently across R.
Supplementary Figure 5. Number of fine-tuning samples required for the PSNR values for
ImageNet-trained networks to converge. Average PSNR values across T2-weighted validation
images were measured for the ImageNet-trained networks trained on (a) 500, (b) 1000, (c) 2000,
and (d) 4000 images. Convergence was taken as the number of fine-tuning samples where the
percentage change in PSNR by incrementing N tune fell below 0.05% of the average PSNR for
the T2-trained network (see Supp. Figure. 4). Domain-transferred networks trained on fewer
samples require more fine-tuning samples for the PSNR values to converge. Furthermore, at
higher values of R, more fine-tuning samples are required for convergence.
Supplementary Table 1. Reconstruction quality for single-coil T1-weighted images
undersampled at R= 2, 4, 6, 8, 10. Reconstructions were performed via ImageNet-trained, T1-
trained and T2-trained networks, as well as conventional CS. PSNR and SSIM values are
reported as mean±standard deviation across test images. Results are shown for raw networks
trained on 2000 training images (raw), and fine-tuned networks tuned with few tens of T1-
weighted images (tuned).

ImageNet-trained T1-trained T2-trained

PSNR SSIM PSNR SSIM PSNR SSIM
Raw 40.33 ± 3.42 0.96 ± 0.02 40.65 ± 3.07 0.97 ± 0.01 40.15 ± 3.14 0.96 ± 0.01
R=2

Tuned 42.81 ± 3.32 0.97 ± 0.01 42.37 ± 3.25 0.97 ± 0.01 42.75 ± 3.22 0.97 ± 0.01
CS
PSNR SSIM
37.54 ± 3.33 0.93 ± 0.24
ImageNet-trained T1-trained T2-trained
PSNR SSIM PSNR SSIM PSNR SSIM
Raw 34.07 ±3.19 0.89 ± 0.03 34.87 ± 2.90 0.91 ± 0.02 33.26 ± 3.23 0.90 ± 0.03
R=4

Tuned 35.85 ± 3.03 0.93± 0.03 36.09 ± 3.19 0.93 ± 0.03 35.95 ±3.03 0.93 ± 0.03
CS
PSNR SSIM
31.77 ± 3.51 0.84 ± 0.04
ImageNet-trained T1-trained T2-trained
PSNR SSIM PSNR SSIM PSNR SSIM
Raw 29.42 ± 3.59 0.84 ± 0.04 32.34 ± 2.95 0.89 ± 0.03 30.48 ± 3.22 0.86 ± 0.03
R=6

Tuned 33.47 ± 3.11 0.90± 0.03 33.90 ± 3.26 0.90 ± 0.04 33.63 ± 3.09 0.90 ± 0.03
CS
PSNR SSIM
29.71 ± 3.52 0.79 ± 0.05
ImageNet-trained T1-trained T2-trained
PSNR SSIM PSNR SSIM PSNR SSIM
Raw 27.28 ± 3.77 0.81 ± 0.04 30.07 ± 3.18 0.86 ± 0.03 28.42 ± 3.14 0.83 ± 0.04
R=8

Tuned 32.14 ± 3.22 0.89 ± 0.04 32.21 ± 3.32 0.89 ± 0.04 32.17 ± 3.45 0.89 ± 0.04
CS
PSNR SSIM
28.56 ± 3.53 0.76 ± 0.06
ImageNet-trained T1-trained T2-trained
PSNR SSIM PSNR SSIM PSNR SSIM
Raw 25.82 ± 3.85 0.79 ± 0.05 28.84 ± 3.43 0.85 ± 0.04 27.72 ± 3.30 0.82 ± 0.04
R=10

Tuned 30.93 ± 3.40 0.87 ± 0.04 31.53 ± 3.38 0.88 ± 0.04 31.42 ± 3.28 0.88 ± 0.04
CS
PSNR SSIM
27.98 ± 3.49 0.75 ± 0.06
Supplementary Table 2. Reconstruction quality for single-coil T2-weighted images
undersampled at R= 2, 4, 6, 8, 10. Reconstructions were performed via ImageNet-trained, T1-
trained and T2-trained networks, as well as conventional CS. PSNR and SSIM values are
reported as mean±standard deviation across test images. Results are shown for raw networks
trained on 2000 training images (raw), and fine-tuned networks tuned with few tens of T2-
weighted images (tuned).

ImageNet-trained T1-trained T2-trained

PSNR SSIM PSNR SSIM PSNR SSIM
Raw 38.84 ± 1.29 0.95 ± 0.01 38.30 ± 1.49 0.94 ± 0.01 39.93 ± 1.48 0.96 ± 0.01
R=2

Tuned 41.81 ± 1.29 0.97 ± 0.01 41.38 ± 1.31 0.97 ± 0.01 41.79 ± 1.29 0.97 ± 0.01
CS
PSNR SSIM
35.94 ± 1.30 0.92 ± 0.01
ImageNet-trained T1-trained T2-trained
PSNR SSIM PSNR SSIM PSNR SSIM
Raw 33.00 ± 1.46 0.88 ± 0.02 32.81 ± 1.60 0.87 ± 0.02 33.94 ± 1.51 0.90 ± 0.02
R=4

Tuned 35.30 ± 1.38 0.92 ± 0.01 35.62 ± 1.40 0.93 ± 0.01 35.45 ± 1.41 0.92 ± 0.01
CS
PSNR SSIM
29.79 ± 1.51 0.81 ± 0.03
ImageNet-trained T1-trained T2-trained
PSNR SSIM PSNR SSIM PSNR SSIM
Raw 30.35 ± 1.38 0.84 ± 0.02 30.90 ± 1.47 0.85 ± 0.03 31.68 ± 1.38 0.87 ± 0.02
R=6

Tuned 33.05 ± 1.36 0.90 ± 0.02 32.98 ± 1.39 0.90 ± 0.02 33.15 ± 1.40 0.90 ± 0.02
CS
PSNR SSIM
27.71 ± 1.54 0.75 ± 0.03
ImageNet-trained T1-trained T2-trained
PSNR SSIM PSNR SSIM PSNR SSIM
Raw 29.03 ± 1.40 0.81 ± 0.02 29.37 ± 1.44 0.82 ± 0.03 30.28± 1.37 0.85 ± 0.02
R=8

Tuned 31.44 ± 1.37 0.87 ± 0.02 31.49 ± 1.39 0.88 ± 0.02 31.70 ± 1.36 0.88 ± 0.02
CS
PSNR SSIM
26.74 ± 1.55 0.72 ± 0.03
ImageNet-trained T1-trained T2-trained
PSNR SSIM PSNR SSIM PSNR SSIM
Raw 27.96 ± 1.40 0.77 ± 0.03 28.29 ± 1.44 0.80 ± 0.03 29.51 ± 1.34 0.84 ± 0.03
R=10

Tuned 30.64 ± 1.35 0.86 ± 0.02 31.01 ± 1.35 0.87 ± 0.02 31.10 ± 1.34 0.87 ± 0.02
CS
PSNR SSIM
26.16 ± 1.54 0.71 ± 0.04