Journal of

Imaging

Article
An Optimization-Based Meta-Learning Model for MRI
Reconstruction with Diverse Dataset
Wanyu Bian 1, *,† , Yunmei Chen 1 , Xiaojing Ye 2 and Qingchao Zhang 1

1 Department of Mathematics, University of Florida, Gainesville, FL 32611, USA; [email protected] (Y.C.);

[email protected] (Q.Z.)
2 Department of Mathematics and Statistics, Georgia State University, Atlanta, GA 30303, USA; [email protected]
* Correspondence: [email protected]
† Author list in alphabetical order.

Abstract: This work aims at developing a generalizable Magnetic Resonance Imaging (MRI) recon-
struction method in the meta-learning framework. Specifically, we develop a deep reconstruction
network induced by a learnable optimization algorithm (LOA) to solve the nonconvex nonsmooth
variational model of MRI image reconstruction. In this model, the nonconvex nonsmooth regular-
ization term is parameterized as a structured deep network where the network parameters can be
learned from data. We partition these network parameters into two parts: a task-invariant part for the
common feature encoder component of the regularization, and a task-specific part to account for the
variations in the heterogeneous training and testing data. We train the regularization parameters in a
bilevel optimization framework which significantly improves the robustness of the training process
and the generalization ability of the network. We conduct a series of numerical experiments using




heterogeneous MRI data sets with various undersampling patterns, ratios, and acquisition settings.

 The experimental results show that our network yields greatly improved reconstruction quality over
Citation: Bian, W.; Chen, Y.; Ye, X.; existing methods and can generalize well to new reconstruction problems whose undersampling
Zhang, Q. An Optimization-Based patterns/trajectories are not present during training.
Meta-Learning Model for MRI
Reconstruction with Diverse Dataset. Keywords: MRI reconstruction; meta-learning; domain generalization
J. Imaging 2021, 7, 231. https://
doi.org/10.3390/jimaging7110231

Academic Editors: Fabiana Zama and 1. Introduction

Elena Loli Piccolomini
Deep learning methods have demonstrated promising performance in a variety of
image reconstruction problems. However, deep learning models are often trained for
Received: 23 September 2021
Accepted: 28 October 2021
specific tasks and require the training samples to follow the corresponding distribution. In
Published: 31 October 2021
particular, the source-domain/training samples and target-domain/testing samples need
to be drawn from the same distribution [1–4]. In practice, these data sets are often collected
Publisher’s Note: MDPI stays neutral
at different sources and exhibit substantial heterogeneity, and thus the samples may follow
with regard to jurisdictional claims in
related but different distributions in real-world applications [5,6]. Therefore, the robust and
published maps and institutional affil- efficient training of deep neural networks using such data sets is theoretically important
iations. and practically relevant in the application of deep learning-based methods.
Meta-learning provides a unique paradigm to achieve robust and efficient neural
network training [1,4,7–10]. Meta-learning is known as learning-to-learn and aims to quickly
learn unseen tasks from the experience of learning episodes that cover the distribution
of relevant tasks. In a multiple-task scenario, given a family of tasks, meta-learning has
Copyright: © 2021 by the authors.
Licensee MDPI, Basel, Switzerland.
been proven to be a useful tool for extracting task-agnostic knowledge and improving the
This article is an open access article
learning performance of new tasks from that family [11,12]. We leverage this feature of
distributed under the terms and meta-learning for network training where the MRI training data are acquired by using
conditions of the Creative Commons different under-sampling patterns (e.g., Cartesian mask, Radial mask, Poisson mask),
Attribution (CC BY) license (https:// under-sampling ratios, and different settings of the scanning parameters, which result in
creativecommons.org/licenses/by/ different levels of contrast (e.g., T1-weighted, T2-weighted, proton-density (PD), and Flair).
4.0/). These data are vastly heterogeneous and can be considered as being from various tasks.

J. Imaging 2021, 7, 231. https://doi.org/10.3390/jimaging7110231 https://www.mdpi.com/journal/jimaging

J. Imaging 2021, 7, 231 2 of 29

Thus, our goal is to develop a robust and generalizable image reconstruction method in
the meta-learning framework to leverage such large-scale heterogeneous data for MRI
reconstruction.
Our approach can be outlined as follows. First, we introduce a variational model
rendering a nonconvex nonsmooth optimization problem for image reconstruction. In our
variational model, the regularization term is parameterized as a structured deep neural
network where the network parameters can be learned during the training process. We then
propose a learnable optimization algorithm (LOA) with rigorous convergence guarantees
to solve this optimization problem. Then, we construct a deep reconstruction network by
following this LOA exactly; namely, each phase of this LOA-induced network is exactly
one iteration of the LOA. This approach is inspired by [13], but the LOA developed in the
present work is computationally more efficient than that in [13]: the safeguard variable in
this work is updated only if necessary, which can significantly reduce computational cost
while retaining the convergence guarantee.
Second, to improve network robustness and mitigate the overfitting issue, we explicitly
partition the network parameters of the regularization into two parts: a task-invariant
part and a task-specific part. The former extracts common prior information of images
from different tasks in the training data and learns the task-invariant component of the
regularization. The latter, on the other hand, is obtained from another network which
exploits proper task-specific components (also called meta-knowledge) of the regularization
for different tasks. The hyperparameters of this network are also learned during training.
Furthermore, we split the available data into two sets: the training set and the validation
set. Then, we introduce a bilevel optimization model for learning network parameters.
Specifically, the lower-level problem (also known as inner problem) finds the task-invariant
part of the regularization with the fixed task-specific part on the training dataset, whereas
the upper-level (outer) problem seeks for the optimal task-specific part of the regularization
parameter using the validation dataset. This approach greatly increases the robustness of
the learned regularization, meaning that the trained LOA-induced deep reconstruction
network can generalize well to unseen tasks.
As demonstrated by our numerical experiments in Section 5, our proposed framework
yields much improved image qualities using diverse data sets of various undersampling
trajectories and ratios for MRI image reconstruction. The reason is that effective regular-
ization can integrate common features and prior information from a variety of training
samples from diverse data sets, but they need to be properly weighed against the data
fidelity term obtained in specific tasks (i.e., undersampling trajectory and ratios). Our
contributions can be summarized as follows:
1. An LOA inspired network architecture—our network architecture exactly follows a
proposed LOA with guaranteed convergence. Thus, the network is more interpretable,
parameter-efficient, and stable than existing unrolling networks.
2. Adaptive design of regularization—our adaptive regularizer consists of a task-invariant
part and a task-specific part, both of which can be appropriately trained from data.
3. Improved network robustness and generalization ability—we improve the robustness
of the network parameter training process by posing it as a bilevel optimization using
training data in the lower-level and validation data in the upper-level. This approach
also improves the generalization ability of the trained network so that it can be quickly
adapted to image reconstruction with new unseen sampling trajectories and produces
high-quality reconstructions.
The remainder of the paper is organized as follows. In Section 2, we discuss related
work for both optimization-based meta-learning and deep unrolled networks for MRI recon-
structions. We propose our meta-learning model and the neural network in Section 3 and
describe the implementation details in Section 4. Section 5 provides the numerical results of
the proposed method. Section 6 concludes the paper.
J. Imaging 2021, 7, 231 3 of 29

2. Related Work
In recent years, meta-learning methods have demonstrated promising results in var-
ious fields with different techniques [12]. Meta-learning techniques can be categorized
into three groups [14–16]: metric-based methods [17–19], model-based methods [20–23],
and optimization-based methods [8,24,25]. Optimization-based methods are often cast as a
bilevel optimization problem and exhibit relatively better generalizability for wider task
distributions. We mainly focus on optimization-based meta-learning in this paper. For
more comprehensive literature reviews and developments of meta-learning, we refer the
readers to the recent surveys [12,16].
Optimization-based meta-learning methods have been widely used in a variety of deep
learning applications [8,24–31]. The network training problems in these meta-learning
methods are often cast as the bilevel optimization of a leader variable and a follower
variable. The constraint of the bilevel optimization is that the follower variable is optimal
for the lower-level problem for each fixed leader variable, and the ultimate goal of bilevel
optimization is to find the optimal leader variable (often, the corresponding optimal
follower variable as well) that minimizes the upper-level objective function under the
constraint. The lower-level problem is approximated by one or a few gradient descent
steps in many existing optimization-based meta learning applications, such as Model-
Agnostic Meta-Learning (MAML) [8], and a large number of followup works of MAML
proposed to improve generalization using similar strategy [9,15,27,29,30,32–34]. Deep
bilevel learning [35] seeks to obtain better generalization than when trained on one task
and generalize well to another task. The model is used to optimize a regularized loss
function to find network parameters from the training set and identify hyperparameters so
that the network performs well on the validation dataset.
When the unseen tasks lie in inconsistent domains with the meta-training tasks, as
revealed in [36], the generalization behavior of the meta-learner will be compromised. This
phenomenon partially arises from the meta-overfitting on the already seen meta-training
tasks, which is identified as a memorization problem in [34]. A meta-regularizer forked with
information theory is proposed in [34] to handle the memorization problem by regulating
the information dependency during the task adaption. MetaReg [4] decouples the entire
network into the feature network and task network, where the meta-regularization term is
only applied to the task network. They first update the parameters of the task network with
a meta-train set to obtain the domain-aligned task network and then update the parameters
of the meta-regularization term on the meta-test set to learn the cross-domain generalization.
In contrast to MetaReg, Feature-Critic Networks [37] exploit the meta-regularization term to
pursue a domain-invariant feature extraction network. The meta-regularization is designed
as a feature-critic network that takes the extracted feature as an input. The parameters
of the feature extraction network are updated by minimizing the new meta-regularized
loss. The auxiliary parameters in the feature-critic network are learned by maximizing the
performance gain over the non-meta case. To effectively evaluate the performance of the
meta-learner, several new benchmarks [38–40] were developed under more realistic settings
that operate well on diverse visual domains. As mentioned in [39], the generalization
to unseen tasks within multimodal or heterogeneous datasets remains a challenge to the
existing meta-learning methods.
The aforementioned methods pursue domain generalization for the classification
networks that learned a regularization function to learn cross-domain generalization.
Our proposed method was developed to solve the inverse problem, and we construct an
adaptive regularization that not only learns the universal parameters among tasks but also
the task-aware parameters. The designated adaptive regularizer assists the generalization
ability of the deep model so that the well-trained model can perform well on heterogeneous
datasets of both seen and unseen tasks.
J. Imaging 2021, 7, 231 4 of 29

3. Proposed Method
3.1. Preliminaries
We first provide the background of compressed sensing MRI (CS-MRI), the image
reconstruction problem, and the learned optimization algorithm to solve the image recon-
struction problem. CS-MRI accelerates MRI data acquisition by under-sampling the k-space
(Fourier space) measurements. The under-sampled k-space measurement are related to the
image by the following formula [41]:

y = PF x + n, (1)

where y ∈ C p represents the measurements in k-space with a total of p sampled data points,
x ∈ C N ×1 is the MR image to be reconstructed with N pixels, F ∈ C N × N is the 2D discrete
Fourier transform (DFT) matrix, and P ∈ R p× N ( p < N ) is the binary matrix representing
the sampling trajectory in k-space. n is the acquisition noise in k-space.
Solving x from (noisy) under-sampled data y according to (1) is an ill-posed problem.
An effective strategy to elevate the ill-posedness issue is to incorporate prior information to
the reconstruction. The variational method is one of the most effective ways to achieve this.
The general framework of this method is to minimize an objective function that consists of
a data fidelity term and a regularization term as follows:

1
x̄ = arg min kPF x − yk2 + R(x), (2)
x 2

where the first term is data fidelity, which ensures consistency between the reconstructed
x and the measured data y, and the second term R(x) is the regularization term, which
introduces prior knowledge to the image to be reconstructed. In traditional variational
methods, R(x) is a hand-crafted function such as Total Variation (TV) [42]. The advances
of the optimization techniques allowed more effective algorithms to solve the variational
models with theoretical justifications. However, hand-crafted regularizers may be too
simple to capture subtle details and satisfy clinic diagnostic quality.
In recent years, we have witnessed the tremendous success of deep learning in solving
a variety of inverse problems, but the interpretation and generalization of these deep-
learning-based methods still remain the main concerns. As an improvement over generic
black-box-type deep neural networks (DNNs), several classes of learnable optimization
algorithms (LOAs) inspired neural networks, known as unrolling networks, which unfold
iterative algorithms to multi-phase networks and have demonstrated promising solution
accuracy and efficiency empirically [43–50]. However, many of them are only specious
imitations of the iterative algorithms and hence lack the backbone of the variational model
and any convergence guarantee.
In light of the substantial success of deep learning and the massive amount of training
data now available, we can parameterize the regularization term as a deep convolutional
neural network (CNN) that learns from training samples. LOA-induced reconstruction
methods have been successfully applied to CS-MRI to solve inverse problems with a
learnable regularizer:
1
arg min kPF x − yk2 + R(x; Θ). (3)
x 2
where R(x; Θ) is the regularization parameterized as a deep network with parameter
Θ. Depending on the specific parametric form of R(x; Θ) and the optimization scheme
used for unrolling, several unrolling networks have been proposed in recent years. For
example, the variational network (VN) [51] was introduced to unroll the gradient descent
algorithm and parametrize the regularization as a combination of linear filters and nonlinear
CNNs. MoDL [52] proposed a weight sharing strategy in a recursive network to learn the
regularization parameters by unrolling the conjugate gradient method. ADMM-Net [53]
mimics the celebrated alternating direction method of multipliers; the regularizer is designed
to be L1 -norm replaced by a piecewise linear function. ISTA-Net [54] considers the regularizer
J. Imaging 2021, 7, 231 5 of 29

as the L1 -norm of a convolutional network. The network unrolls several phases iteratively, and
each phase mimics one iteration of iterative shrinkage thresholding algorithm (ISTA) [55,56].
However, these networks only superficially mimic the corresponding optimization schemes,
but they lack direct relations to the original optimization method or variational model
and do not retain any convergence guarantee. In this work, we first develop a learnable
optimization algorithm (LOA) for (3) with comprehensive convergence analysis and obtain
an LOA-induced network by following the iterative scheme of the LOA exactly.

3.2. LOA-Induced Reconstruction Network

In this section, we first introduce a learned optimization algorithm (LOA) to solve (3)
where the regularization network parameter Θ is fixed. As Θ is fixed in (3), we temporarily
omit this in the derivation of the LOA below and write R(x; Θ) as R(x) for notation
simplicity.
In this work, to incorporate sparsity along with the learned features, we parameterize
the function R(x) = κ · r (x), where κ > 0 is a weight parameter that needs be chosen
properly depending on the specific task (e.g., noise level, undersampling ratio, etc.), and r is a
regularizer parameterized as a composition of neural networks and can be adapted to a broad
range of imaging applications. Specifically, we parameterize r as the composition of the l2,1
norm and a learnable feature extraction operator g(x). That is, we set r in (11) to be
m
r (x) := kg(x)k2,1 = ∑ kgj (x)k. (4)
j =1

Here, “:=” stands for “defined as”. g(·) = (g1 (·), . . . , gm (·)), g j (·) = g j (·; θ ) is parametrized
as a convolutional neural network (CNN) for j = 1, · · · , m, and θ is the learned and fixed
network parameter in r (·; θ ), as mentioned above. We also consider κ to be learned and
fixed as θ for now, and we discuss how to learn both of them in the next subsection. We
use a smooth activation function in g as formulated in (20), which renders g a smooth but
nonconvex function. Due to the nonsmooth k · k2,1 , r is therefore a nonsmooth nonconvex
function.
Since the minimization problem in (11) is nonconvex and nonsmooth, we need to
derive an efficient LOA to solve it. Here, we first consider smoothing the l2,1 norm that for
any fixed g(x) is q
m
r ε (x) = ∑ j =1 kg j (x)k2 + ε2 − ε. (5)

We denote Rε = κ · rε . The LOA derived here is inspired by the proximal gradient descent
algorithm and iterates the following steps to solve the smoothed problem:

zt +1 = xt − α t ∇ f (xt ) (6a)
xt+1 = proxαt Rε (zt+1 ), (6b)
t

where ε t denotes the smoothing parameter ε at the specific iteration t, and the proximal
operator is defined as proxαg (b) := arg minx kx − bk + αg(x) in (6b). A quick observation
from (5) is that Rε → R as ε diminishes, so later we intentionally push ε t → 0 at Line 16
in Algorithm 1. Then, one can readily show that Rε ( x ) ≤ R( x ) ≤ Rε ( x ) + ε for all x and
ε > 0. From Algorithm 1, line 16 automatically reduces ε, and the iterates will converge to
the solution of the original nonsmooth nonconvex problem (11)—this is clarified precisely
in the convergence analysis in Appendix A.
Since Rε t is a complex function involving a deep neural network, its proximal operator
does not have a closed form and cannot be computed easily in the subproblem in (6b). To
overcome this difficulty, we consider to approximate Rε t by

1
R̂ε t (zt+1 ) = Rε t (zt+1 ) + h∇ Rε t (zt+1 ), x − zt+1 i + k x − zt +1 k 2 . (7)
2β t
J. Imaging 2021, 7, 231 6 of 29

Then, we update ut+1 = proxαt R̂ε (zt+1 ) to replace (6b); therefore, we obtain
t

αt β t
ut+1 = zt+1 − τt ∇ Rε t (zt+1 ), where τt = . (8)
αt + β t

In order to guarantee the convergence of the algorithm, we introduce the standard gradient
descent of φε t (where φε t := f + Rε t ) at x:

1
vt+1 = arg minh∇ f (xt ), x − xt i + h∇ Rε (xt ), x − xt i + k x − xt k 2 , (9)
x 2αt

which yields
vt+1 = xt − αt ∇φε t (xt ), (10)
to serve as a safeguard for the convergence. Specifically, we set xt+1 = ut+1 if φε t (ut+1 ) ≤
φε t (vt+1 ); otherwise, we set xt+1 = vt+1 . Then, we repeat this process.
Our algorithm is summarized in Algorithm 1. The prior term with unknown pa-
rameters has the exact residual update itself which improves the learning and training
process [57]. The condition checking in Line 5 is introduced to make sure that it is in the
energy descending direction. Once the condition in Line 5 fails, the process moves to vt+1 ,
and the line search in Line 12 guarantees that the appropriate step size can be achieved
within finite steps to make the function value decrease. From Line 3 to Line 14, we consider
that it solves a problem of minimizing φε t with ε t fixed. Line 15 is used to update the value
of ε t depending on a reduction criterion. The detailed analysis of this mechanism and
in-depth convergence justification is shown in Appendix A. The corresponding unrolling
network exactly follows Algorithm 1 and thus shares the same convergence property. Com-
pared to LDA [13], which computes both candidates ut+1 , vt+1 at every iteration and then
chooses the one that achieves a smaller function value, we propose the criteria above in
Line 5 for updating xt+1 , which potentially saves extra computational time for calculating
the candidate vt+1 and potentially mitigates the frequent alternations between the two
candidates. Besides, the smoothing method proposed in this work is more straightforward
than smoothing in dual space [13] while still keeping provable convergence, as shown in
Theorem A5.
The proposed LOA-induced network is a multi-phase network whose architecture
exactly follows the proposed LOA (Algoirthm 1) in the way that each phase corresponds
to one iteration of the algorithm. Specifically, we construct a deep network, denoted by
FΘ , that follows Algorithm 1 exactly for a user-specified number of T iterations. Here, Θ
denotes the set of learnable parameters in FΘ , which includes the regularization network
parameter θ, weight κ, and other algorithmic parameters of Algorithm 1. Therefore, for
any input under-sampled k-space measurement y, FΘ (y) executes the LOA (Algorithm 1)
for T iterations and generates an approximate solution of the minimization problem (11):

FΘ (y) ≈ arg min φΘ (x, y) := f (x, y) + R(x; Θ) .


(11)
x

where we use “≈” since FΘ follows only finitely many steps of the optimization algorithm
to approximate the solution. It is worth emphasizing that this approach can be readily
applied to a much broader class of image reconstruction problems as long as f is (possibly
nonconvex and) continuously differentiable with the Lipschitz continuous gradient. In the
next subsection, we develop a meta-learning based approach for the robust training of the
network parameter Θ.

3.3. Bilevel Optimization Algorithm for Network Training

In this section, we consider the parameter training problem of the LOA-induced
network FΘ . Specifically, we develop a bilevel optimization algorithm to train our network
parameters Θ from diverse data sets to improve network robustness and generalization
ability.
J. Imaging 2021, 7, 231 7 of 29

Algorithm 1: Algorithmic Unrolling Method with Provable Convergence

1: Input: Initial x0 , 0 < ρ, γ < 1, and ε 0 , a, σ > 0. Max total phases T or tolerance
etol > 0.
2: for t = 0, 1, 2, . . . , T − 1 do
3: zt +1 = xt − α t ∇ f (xt )
4: ut+1 = zt+1 − τt ∇ Rε t (zt+1 ),
5: if k∇φε t (xt )k ≤ akut+1 − xt k and φε t (ut+1 ) − φε t (xt ) ≤ − 1a kut+1 − xt k2 then
6: set xt+1 = ut+1 ,
7: else
8: vt+1 = xt − αt ∇φε t (xt ),
9: if φε t (vt+1 ) − φε t (xt ) ≤ − 1a kvt+1 − xt k2 holds then
10: set xt+1 = vt+1 ,
11: else
12: update αt ← ραt , then go to 8,
13: end if
14: end if
15: if k∇φε t (xt+1 )k < σγε t , set ε t+1 = γε t ; otherwise, set ε t+1 = ε t .
16: if σε t < etol , terminate.
17: end for
18: Output: xt .

Recall that the LOA-induced network FΘ exactly follows Algorithm 1, which is de-
signed to solve the variational model (11) containing learnable regularization R(x; Θ). As
shown in Section 3.2, we design R(x; Θ) = κ · r (x; Θ), where r is learned to capture the
intrinsic property of the underlying common features across all different tasks. To account
for the large variations in the diverse training/validation data sets, we introduce a task-
specific parameter ωi to approximate the proper κ for the ith task. Specifically, for the ith
task, the weight κ is set to σ (ωi ) ∈ (0, 1), where σ (·) is the sigmoid function. Therefore,
κ = σ (ωi ) finds the proper weight of r for the i-th task according to its specific sampling
ratio or pattern. The parameters ωi are to be optimized in conjunction with Θ through the
hyperparameter tuning process below.
Suppose that we are given M data pairs {(ym , x∗m )}M m=1 for the use of training and
validation, where ym is the observation, which is the partial k-space data in our setting,
and x∗m is the corresponding ground truth image. The data pairs are then sampled into
N tasks {Dτi }iN=1 , where each Dτi represents the collection of data pairs in the specific
task τi . In each task τi , we further divide the data into the task-specific training set Dτtri
and validation set Dτvali
. The architecture of our base network exactly follows the LOA
(Algorithm 1) developed in the previous section with learnable parameters θ and a task-
specific parameter ωi for the ith task. More precisely, for one data sample denoted by
(i ) (i )
(y j , x∗ j ) in task τi with index j, we propose the algorithmic unrolling network for task τi
as
(i ) (i )
Fθ,ωi (y j ) ≈ arg min f (x, y j ) + σ (ωi )r (x; θ ), (12)
x

where θ denotes the learnable common parameters across different tasks with task-invariant
representation, whereas ωi is a task-specific parameter for task τi . The weight σ (ωi )
represents the weight of r associated with the specific task τi . In our proposed network,
Θ is the collection of (θ, ωi ) for task i = 1, · · · N . We denote ω to be the set {ωi }iN=1 . The
detailed architecture of this network is illustrated in Section 3.2. We define the task-specific
loss
|Dτi |
`τi (θ, ωi ; Dτi ) := ∑ ` Fθ,ωi (y j ), x ∗ j ,
(i ) (i )

(13)
j =1
J. Imaging 2021, 7, 231 8 of 29

where |Dτi | represents the cardinality of Dτi and

(i ) (i ) 1 (i ) (i )
k F (y ) − x ∗ j k2 .


` Fθ,ωi (y j ), x ∗ j := (14)
2 θ,ωi j
For the sake of preventing the proposed model from overfitting the training data, we
introduce a novel learning framework by formulating the network training as a bilevel
optimization problem to learn ω and θ in (12) as

N
min
θ, ω ={ωi :i ∈[ N ]}
∑ `τi (θ (ω ), ωi ; Dτvali ) (15a)
i =1
N
s.t. θ (ω ) = arg min ∑ `τi (θ, ωi ; Dτtri ). (15b)
θ i =1

In (15), the lower-level optimization learns the task-invariant parameters θ of the feature
encoder with the fixed task-specific parameter ωi on the training dataset, and the upper-
level adjusts the task-specific parameters {ωi } so that the task-invariant parameters θ can
perform robustly on the validation dataset as well. For simplicity, we omit the summation
and redefine L(θ, ω; D) := ∑iN=1 `τi (θ, ω; D) and then briefly rewrite (15) as

min L(θ (ω ), ω; D val ) s.t. θ (ω ) = arg min L(θ, ω; D tr ). (16)

θ,ω θ

Then, we relax (16) into a single-level constrained optimization where the lower-level
problem is replaced with its first-order necessary condition following [58]

min L(θ, ω; D val ) s.t. ∇θ L(θ, ω; D tr ) = 0. (17)

θ,ω

which can be further approximated by an unconstrained problem by a penalty term as

e θ, ω; D tr , D val ) := L(θ, ω; D val ) + λ k∇θ L(θ, ω; D tr )k2 .


min L( (18)
θ,ω 2

We adopt the stochastic gradients of the loss functions on mini-batch data sets in
each iteration. In our model, we need to include the data pairs of multiple tasks in one
batch; therefore, we propose the cross-task mini-batches when training. At each training
(i ) (i ) tr
iteration, we randomly sample the training data pairs Bτtri = {(y j , x∗ j ) ∈ Dτtri }J
j=1 and
(i ) (i ) val
the validation pairs Bτval
i
= {(y j , x∗ j ) ∈ Dτval
i
}Jj=1 on each task τi . Then, the overall
training and validation mini-batches B tr and B val used in every training iteration are
composed of the sampled data pairs from the entire set of tasks; i.e., B tr = iN=1 {Bτtri } and
S

B val = iN=1 {Bτval }. Thus in each iteration, we have N · J tr and N · J val data pairs used
S
i
for training and validation, respectively. To solve the minimization problem (17), we utilize
the stochastic mini-batch alternating direction method summarized in Algorithm 2, which
is modified from [58].
As analyzed in [58], this penalty-type method has linear time complexity without
computing the Hessian of the low level and only requires a constant space since we
only need to store the intermediate θ, ω at each training iteration, which is suitable for
solving the large-scale bilevel optimization problem. Algorithm 2 relaxes the bi-level
optimization problem to a single-level constrained optimization problem by using the
first-order necessary condition, which is not equivalent to the original problem but is much
easier and efficient to solve. In the inner-loop (Line 5–9) of Algorithm 2, we continue
minimizing the converted single-level optimization function (18) with respect to θ for K
steps and then ω once alternatively until the condition with tolerance δ in Line 5 fails. The
basic idea behind the condition in Line 5 arises from the first-order necessary condition
as we would like to push the gradient of Le toward 0. Furthermore, at Line 11 of the outer
J. Imaging 2021, 7, 231 9 of 29

loop (Line 2–11), we decrease the tolerance δ. Combining Line 5 and 11 guarantees that
each time the inner loop terminates, the gradients of Le with respect to θ and ω become
increasingly close to 0. The parameter δtol is used to control the accuracy of the entire
algorithm, and the outer-loop will terminate when δ is sufficiently small (i.e., δ ≤ δtol ). In
addition, λ is the weight for the second constraint term of (18); in the beginning, we set λ
to be small to achieve a quick starting convergence, then gradually increase its value to
emphasize the constraint.

Algorithm 2: Stochastic mini-batch alternating direction penalty method to solve

problem (17)
1: Input Dτtri , Dτval
i
, δtol > 0.
2: Initialize θ, ω, δ, λ > 0 and νδ ∈ (0, 1), νλ > 1.
3: while δ > δtol do
(i ) (i )
Sample cross-task training batch B tr = iN=1 {(y j , x∗ j ) ∈ Dτtri } j=1:J tr
S
4:
(i ) (i )
Sample cross-task validation batch B val = iN=1 {(y j , x∗ j ) ∈ Dτval } j=1:J val
S
5:
i

6: while k∇θ L( tr val

e θ, ω; B , B )k + k∇ω L(2 tr val 2
e θ, ω; B , B )k > δ do
7: for k = 1, 2, . . . , K (inner loop) do
8: θ ← θ − ρkθ ∇θ L( e θ, ω; B tr , B val )
9: end for
10: ω ← ω − ρω ∇ω L( e θ, ω; B tr , B val )
11: end while
12: update δ ← νδ δ, λ ← νλ λ
13: end while
14: output: θ, ω.

4. Implementation
4.1. Feature Extraction Operator
We set the feature extraction operator g to be a vanilla l-layer CNN with the component-
wise nonlinear activation function ϕ and no bias, as follows:

g( x ) = wl ∗ ϕ · · · ϕ(w3 ∗ ϕ(w2 ∗ ϕ(w1 ∗ x ))), (19)

where {wq }lq=1 denote the convolution weights consisting of d kernels with identical spatial
kernel size, and ∗ denotes the convolution operation. Here, ϕ is constructed to be the
smoothed rectified linear unit as defined below:

 0,
 if x ≤ −δ,
1 2 1
4δ x + 2 x + 4 , if − δ < x < δ,
ϕ( x ) = δ (20)

x, if x ≥ δ,


where the prefixed parameter δ is set to be 0.001 in our experiment. The default configura-
tion of the feature extraction operator is set as follows: the feature extraction operator g
consists of l = 3 convolution layers and all convolutions are with 4 kernels of a spatial size
of 3 × 3.

4.2. Setups
As our method introduces an algorithmic unrolling network, there exists a one-to-one
correspondence between the algorithm iterations and the neural network phases (or blocks).
Each phase of the forward propagation can be viewed as one algorithm iteration, which
motivates us to imitate the iterating of the optimization algorithm and use a stair training
strategy [13]. At the first stage, we start training the network parameters using one phase,
then after the the loss converges, we add more phases (one phase each time) then continue
the training process. We repeat this procedure and stop it when the loss does not decrease
J. Imaging 2021, 7, 231 10 of 29

any further when we add more blocks. We minimize the loss for 100 epochs/iterations
each time using the SGD-based optimizer Adam [59] with β 1 = 0.9, β 2 = 0.999, and the
initial learning rate set to 10−3 as well as a mini-batch size of 8. The Xavier Initializer [60] is
used to initialize the weights of all convolutions. The initial smoothing parameter ε 0 is set
to be 0.001 and then learned together with other network parameters. The input x0 of the
unrolling network is obtained by the zero-filling strategy [61]. The deep unrolling network
was implemented using the Tensorflow toolbox [62] in the Python programming language.

5. Numerical Experiments
5.1. Dataset
To validate the performance of the proposed method, the data we used were from
Multimodal Brain Tumor Segmentation Challenge 2018 [63], in which the training dataset
contains four modalities (T1, T1c , T2 and FLAIR )scanned from 285 patients and the
validation dataset contains images from 66 patients, each with a volume size of 240 × 240 ×
155. Each modality consists of two types of gliomas: 75 volumes of low-grade gliomas
(LGG) and 210 volumes of high-grade gliomas (HGG). Our implementation involved HGG
MRI in two modalities—T1 and T2 images—and we chose 30 patients from each modality
in the training dataset to train our network. In the validation dataset, we randomly picked
15 patients as our validation data and 6 patients in the training dataset as testing data,
which were distinct from our training set and validation set. We cropped the 2D image
size to be 160 × 180 in the center region and picked 10 adjacent slices in the center of each
volume, resulting in a total of 300 images as our training data, 150 images as our validation
data, and a total of 60 images as testing data. The amount of data mentioned here is for a
single task, but since we emploedy multi-task training, the total number of images in each
dataset should be multiplied by the number of tasks. For each 2D slice, we normalized the
spatial intensity by dividing the maximum pixel value.

5.2. Experiment Settings

All the experiments were implemented on a Windows workstation with an Intel Core
i9 CPU at 3.3GHz and an Nvidia GTX-1080Ti GPU with 11 GB of graphics card memory via
TensorFlow [64]. The parameters in the proposed network were initialized by using Xavier
initialization [65]. We trained the meta-learning network with four tasks synergistically
associated with four different CS ratios—10%, 20%, 30%, and 40%—and tested the well-
trained model on the testing dataset with the same masks of these four ratios. We used 300
training data for each CS ratio, amounting to a total of 1200 images in the training dataset.
The results for T1 and T2 MR reconstructions are shown in Tables 1 and 2, respectively.
The associated reconstructed images are displayed in Figures 1 and 2. We also tested the
well-trained meta-learning model on unseen tasks with radial masks for unseen ratios
of 15%, 25%, and 35% and random Cartesian masks with ratios of 10%, 20%, 30%, and
40%. The task-specific parameters for the unseen tasks were retrained for different masks
with different sampling ratios individually with fixed task-invariant parameters θ. In this
experiments, we only needed to learn ωi for three unseen CS ratios with radial mask and
four regular CS ratios with Cartesian masks. The experimental training proceeded with
fewer data and iterations, where we used 100 MR images with 50 epochs. For example, to
reconstruct MR images with a CS ratio of 15% from the radial mask, we fixed the parameter
θ and retrained the task-specific parameter ω on 100 raw data points with 50 epochs, then
tested with renewed ω on our testing data set with raw measurements sampled from the
radial mask with a CS radial of 15%. The results associated with radial masks are shown in
Tables 3 and 4, Figures 3 and 4 for T1 and T2 images, respectively. The results associated
with Cartesian masks are listed in Table 5 and reconstructed images are displayed in
Figure 5.
We compared our proposed meta-learning method with conventional supervised
learning, which was trained with one task at each time and only learned the task-invariant
parameter θ without the task-specific parameter ωi . The forward network of conventional
J. Imaging 2021, 7, 231 11 of 29

learning unrolled Algorithm 1 with 11 phases, which was the same as meta-learning. We
merged the training set and validation set, resulting in 450 images for the training of the
conventional supervised learning. The training batch size was set as 25 and we applied a
total of 2000 epochs, while in meta-learning, we applied 100 epochs with a batch size of 8.
The same testing set was used in both meta-learning and conventional learning to evaluate
the performance of these two methods.
We made comparisons between meta-learning and the conventional network on the
seven different CS ratios (10%, 20%, 30%, 40%, 15%, 25%, and 35%) in terms of two types of
random under-sampling patterns: radial sampling mask and Cartesian sampling mask. The
parameters for both meta-learning and conventional learning networks were trained via the
Adam optimizer [59], and they both learned the forward unrolled task-invariant parameter
θ. The network training of the conventional method used the same network configuration as
the meta-learning network in terms of the number of convolutions, depth and size of CNN
kernels, phase numbers and parameter initializer, etc. The major difference in the training
process between these two methods is that meta-learning is performed for multi-tasks
by leveraging the task-specific parameter ωi learned from Algorithm 2, and the common
features among tasks are learned from the feed-forward network that unrolls Algorithm 1,
while conventional learning solves the task-specific problem by simply unrolling the
forward network via Algorithm 1, where both training and testing are implemented on the
same task. To investigate the generalizability of meta-learning, we tested the well-trained
meta-learning model on MR images in different distributions in terms of two types of
sampling masks with various trajectories. The training and testing of conventional learning
were applied with the same CS ratios; that is, if the conventional method was trained with
a CS ratio 10%, then it was also tested on a dataset with a CS ratio of 10%, etc.
Because MR images are represented as complex values, we applied complex convo-
lutions [66] for each CNN; that is, every kernel consisted of a real part and imaginary
part. Three convolutions were used in g, where each convolution contained four filters
with a spatial kernel size of 3 × 3. In Algorithm 1, a total of 11 phases can be achieved if
we set the termination condition etol = 1 × 10−3 , and the parameters of each phase are
shared except for the step sizes. For the hyperparameters in Algorithm 1, we chose an
initial learnable step size α0 = 0.01, τ0 = 0.01, ε 0 = 0.001, and we set prefixed values of
a = 105 , σ = 103 , ρ = 0.9, and γ = 0.9. The principle behind the choices of those parameters
is based on the convergence of the algorithm and effectiveness of the computation. The
parameter 0 < ρ < 1 is the reduction rate of the step size during the line search used to
guarantee the convergence. The parameter 0 < γ < 1 at step 15 is the reduction rate for ε.
In Algorithm 1, from step 2 to step 14, the smoothing level ε is fixed. When the gradient
of the smoothed function is small enough, we reduce ε by a fraction factor γ to find an
approximate accumulation point of the original nonsmooth nonconvex problem. We chose
a larger a in order to have more iterations k for which uk+1 satisfies the conditions in step 5,
so that there would be fewer iterations requiring the computation of vk+1 . Moreover, the
scheme for computing uk+1 is in accordance with the residual learning architecture that
has been proven effective for reducing training error.
In Algorithm 2, we set νδ = 0.95 and the parameter δ was initialized as δ0 = 1 × 10−3
and stopped at value δtol = 4.35 × 10−6 , and a total of 100 epochs were performed. To
train the conventional method, we set 2000 epochs with the same number of phases,
convolutions, and kernel sizes as used to train the meta-learning approach. The initial λ
was set as 1 × 10−5 and νλ = 1.001.
We evaluated our reconstruction results on the testing data sets using three metrics:
peak signal-to-noise ratio (PSNR) [67], structural similarity (SSIM) [68], and normalized
mean squared error (NMSE) [69]. The following formulations compute the PSNR, SSIM,
and NMSE between the reconstructed image x and ground truth x∗ . PSNR can be induced
by the mean square error (MSE) where

max(|x∗ |)

PSNR(x, x∗ ) = 20 log10 p , (21)
MSE(x, x∗ )
J. Imaging 2021, 7, 231 12 of 29

where N is the total number of pixels of the ground truth and MSE is defined by MSE
(x, x∗ ) = N1 kx∗ − xk2 .

(2µx µx∗ + C1 )(2σxx∗ + C2 )

SSI M (x, x∗ ) = , (22)
(µ2x + µ2x∗ + C1 )(σx2 + σx2∗ + C2 )

where µx , µx∗ represent local means, σx , σx∗ denote standard deviations, σxx∗ represents the
covariance between x and x∗ , C1 = (k1 L)2 , C2 = (k2 L)2 are two constants which avoid the
zero denominator, and k1 = 0.01, k2 = 0.03. L is the largest pixel value of MR image.

kx − x∗ k22
N MSE(x, x∗ ) = , (23)
kxk22

where NMSE is used to measure the mean relative error. For detailed information of these
three metrics mentioned above, please refer to [67–69].

5.3. Experimental Results with Different CS Ratios in Radial Mask

In this section, we evaluate the performance of well-trained meta-learning and con-
ventional learning approaches. Tables 1, 2 and 5 report the quantitative results of averaged
numerical performance with standard deviations and associated descaled task-specific
meta-knowledge σ (ωi ). From the experiments implemented with radial masks, we observe
that the average PSNR value of meta-learning improved by 1.54 dB in the T1 brain image
for all four CS ratios compared with the conventional method, and for the T2 brain image,
the average PSNR of meta-learning improved by 1.46 dB. Since the general setting of meta-
learning aims to take advantage of the information provided from each individual task,
with each task associated with an individual sampling mask that may have complemented
sampled points, the performance of the reconstruction from each task benefits from other
tasks. Smaller CS ratios will inhibit the reconstruction accuracy, due to the sparse under-
sampled trajectory in raw measurement, while meta-learning exhibits a favorable potential
ability to solve this issue even in the situation of insufficient amounts of training data.
In general supervised learning, training data need to be in the same or a similar distri-
bution; heterogeneous data exhibit different structural variations of features, which hinder
CNNs from extracting features efficiently. In our experiments, raw measurements sampled
from different ratios of compressed sensing display different levels of incompleteness;
these undersampled measurements do not fall in the same distribution but they are related.
Different sampling masks are shown at the bottom of Figures 1 and 3, and these may have
complemented sampled points, in the sense that some of the points which a 40% sampling
ratio mask did not sample were captured by other masks. In our experiment, different sam-
pling masks provided their own information from their sampled points, meaning that four
reconstruction tasks helped each other to achieve an efficient performance. Therefore, this
explains why meta-learning is still superior to conventional learning when the sampling
ratio is large.
Meta-learning expands a new paradigm for supervised learning—the purpose is to
quickly learn multiple tasks. Meta-learning only learns task-invariant parameters once for
a common feature that can be shared with four different tasks, and each σ (ωi ) provides
task-specific weighting parameters according to the principle of “learning to learn”. In
conventional learning, the network parameter needs to be trained four times with four
different masks since the task-invariant parameter cannot be generalized to other tasks,
which is time-intensive. From Tables 1 and 2, we observe that a small CS ratio needs a
higher value of σ (ωi ). In fact, in our model (11), the task-specific parameters behave as
weighted constraints for task-specific regularizers, and the tables indicate that lower CS
ratios require larger weights to be applied for the regularization.
A qualitative comparison between conventional and meta-learning methods is shown
in Figures 1 and 2, displaying the reconstructed MR images of the same slice for T1 and T2,
respectively. We label the zoomed-in details of HGG in the red boxes. We observe evidence
J. Imaging 2021, 7, 231 13 of 29

that conventional learning is more blurry and loses sharp edges, especially with lower CS
ratios. From the point-wise error map, we find that meta-learning has the ability to reduce
noises, especially in some detailed and complicated regions, compared to conventional
learning.
We also tested the performance of meta-learning with two-thirds of the training and
validation data for the T1-weighted image used in the previous experiment, denoted as
“meta-learning”. For conventional learning, the network was also trained by using two-
thirds of the training samples in the previous experiment. The testing dataset remained
the same as before. These results are displayed in Table 1, where we denote the reduced
data experiments as “meta-learning∗ ” and “conventional∗ ”. These experiments reveal
that the accuracy of test data decreases when we reduce the training data size, but it is
not a surprise that meta-learnining∗ still outperforms conventional learning∗ , and even
conventional learning.
To verify the reconstruction performance of the proposed LOA 1, we compared the
proposed conventional learning with ISTA-Net+ [54], which is a state-of-the-art deep
unfolded network for MRI reconstruction. We retrained ISTA-Net+ with the same training
dataset and testing dataset as conventional learning on the T1-weighted image. For a fair
comparison, we used the same number of convolution kernels, the same dimension of
kernels for each convolution during training, and the same phase numbers as conventional
learning. The testing numerical results are listed in Table 1 and the MRI reconstructions
are displayed in Figure 1. We can observe that the conventional learning which unrolls
Algorithm 1 outperforms ISTA-Net+ in any of the CS ratios. From the corresponding
point-wise absolute error, the conventional learning attains a much lower error and much
better reconstruction quality.

5.4. Experimental Results with Different Unseen CS Ratios in Different Sampling Patterns
In this section, we test the generalizability of the proposed model for unseen tasks. We
fixed the well-trained task-invariant parameter θ and only trained ωi for sampling ratios of
15%, 25%, and 35% with radial masks and sampling ratios of 10%, 20%, 30%, and 40% with
Cartesian masks. In this experiment, we only used 100 training data points for each CS ratio
and applied a total of 50 epochs. The averaged evaluation values and standard deviations
are listed in Tables 3 and 4 for reconstructed T1 and T2 brain images, respectively, with
radial masks, and Table 5 shows the qualitative performance for the reconstructed T2 brain
image with random Cartesian sampling masks applied. In the T1 image reconstruction
results, meta-learning showed an improvement of 1.6921 dB in PSNR for the 15% CS
ratio, 1.6608 dB for the 25% CS ratio, and 0.5764 dB for the 35% ratio compared to the
conventional method, showing the tendency that the level of reconstruction quality for
lower CS ratios improved more than higher CS ratios. A similar trend was found for T2
reconstruction results with different sampling masks. The qualitative comparisons are
illustrated in Figures 3–5 for T1 and T2 images tested with unseen CS ratios in radial masks
and T2 images tested with Cartesian masks with regular CS ratios, respectively. In the
experiments conducted with radial masks, meta-learning was superior to conventional
learning, especially at a CS ratio of 15%—one can observe that the detailed regions in red
boxes maintained their edges and were closer to the true image, while the conventional
method reconstructions are hazier and lost details in some complicated tissues. The point-
wise error map also indicates that meta-learning has the ability to suppress noises.
Training with Cartesian masks is more difficult than radial masks, especially for
conventional learning, where the network is not very deep since the network only applies
three convolutions each with four kernels. Table 5 indicates that the average performance
of Meta-learning improved about 1.87 dB compared to conventional methods with T2 brain
images. These results further demonstrate that meta-learning has the benefit of parameter
efficiency, and the performance is much better than conventional learning even if we apply
a shallow network with a small amount of training data.
J. Imaging 2021, 7, 231 14 of 29

The numerical experimental results discussed above show that meta-learning is capa-
ble of fast adaption to new tasks and has more robust generalizability for a broad range of
tasks with heterogeneous, diverse data. Meta-learning can be considered as an efficient
technique for solving difficult tasks by leveraging the features extracted from easier tasks.

Table 1. Quantitative evaluations of the reconstructions of T1 brain image associated with various
sampling ratios of radial masks. Conventional∗ and meta-learning∗ are trained with two-thirds of
the dataset used in training conventional and meta-learning approaches, respectively.

CS Ratio Methods PSNR SSIM NMSE σ (ωi )

ISTA-Net+ [54] 21.2633 ± 1.0317 0.5487 ± 0.0440 0.1676 ± 0.0253
Conventional∗ 21.6947 ± 1.0264 0.5689 ± 0.0404 0.1595 ± 0.0240
10% Conventional 21.7570 ± 1.0677 0.5650 ± 0.0412 0.0259 ± 0.0082
Meta-learning∗ 22.9633 ± 1.0969 0.5962 ± 0.0415 0.0194 ± 0.0065 0.9339
Meta-learning 23.2672 ± 1.1229 0.6101 ± 0.0436 0.0184 ± 0.0067 0.9218
ISTA-Net+ [54] 26.2734 ± 1.0115 0.7068 ± 0.0364 0.0944 ± 0.0155
Conventional∗ 26.4639 ± 1.0233 0.7107 ± 0.0357 0.0924 ± 0.0154
20% Conventional 26.6202 ± 1.1662 0.7121 ± 0.0397 0.0910 ± 0.0169
Meta-learning∗ 27.9381 ± 1.1121 0.7541 ± 0.0360 0.0063 ± 0.0023 0.8150
Meta-learning 28.2944 ± 1.2119 0.7640 ± 0.0377 0.0058 ± 0.0022 0.7756
ISTA-Net+ [54] 28.8309 ± 1.3137 0.7492 ± 0.0407 0.0708 ± 0.0142
Conventional∗ 29.2923 ± 1.3194 0.7522 ± 0.0399 0.0671 ± 0.0136
30% Conventional 29.5034 ± 1.4446 0.7557 ± 0.0408 0.0657 ± 0.0143
Meta-learning∗ 30.8691 ± 1.5897 0.8310 ± 0.0394 0.0033 ± 0.0015 0.6359
Meta-learning 31.1417 ± 1.5866 0.8363 ± 0.0385 0.0031 ± 0.0014 0.6501
ISTA-Net+ [54] 30.7282 ± 1.5482 0.8008 ± 0.0428 0.0572 ± 0.0127
Conventional∗ 31.3761 ± 1.5892 0.8035 ± 0.0420 0.0532 ± 0.0121
40% Conventional 31.4672 ± 1.6390 0.8111 ± 0.0422 0.0029 ± 0.0014
Meta-learning∗ 32.7330 ± 1.6386 0.8623 ± 0.0358 0.0022 ± 0.0010 0.6639
Meta-learning 32.8238 ± 1.7039 0.8659 ± 0.0370 0.0022 ± 0.0010 0.6447

Table 2. Quantitative evaluations of the reconstructions of T2 brain image associated with various
sampling ratios of radial masks.

CS Ratio Methods PSNR SSIM NMSE σ (ωi )

10% Conventional 23.0706 ± 1.2469 0.5963 ± 0.0349 0.2158 ± 0.0347
Meta-learning 24.0842 ± 1.3863 0.6187 ± 0.0380 0.0112 ± 0.0117 0.9013
20% Conventional 27.0437 ± 1.0613 0.6867 ± 0.0261 0.1364 ± 0.0213
Meta-learning 28.9118 ± 1.0717 0.7843 ± 0.0240 0.0122 ± 0.0030 0.8742
30% Conventional 29.5533 ± 1.0927 0.7565 ± 0.0265 0.1023 ± 0.0166
Meta-learning 31.4096 ± 0.9814 0.8488 ± 0.0217 0.0069 ± 0.0019 0.8029
40% Conventional 32.0153 ± 0.9402 0.8139 ± 0.0238 0.0770 ± 0.0128
Meta-learning 33.1114 ± 1.0189 0.8802 ± 0.0210 0.0047 ± 0.0015 0.7151
J. Imaging 2021, 7, 231 15 of 29

Table 3. Quantitative evaluations of the reconstructions of T1 brain image associated with various
sampling ratios of radial masks. Meta-learning was trained with CS ratios of 10%, 20%, 30%, and
40% and tested with unseen ratios of 15%, 25%, and 35%. The conventional method was subjected to
regular training and testing with the same CS ratios of 15%, 25%, and 35%.

CS Ratio Methods PSNR SSIM NMSE σ (ωi )

15% Conventional 24.6573 ± 1.0244 0.6339 ± 0.0382 0.1136 ± 0.0186
Meta-learning 26.3494 ± 1.0102 0.7088 ± 0.0352 0.0090 ± 0.0030 0.9429
25% Conventional 28.4156 ± 1.2361 0.7533 ± 0.0368 0.0741 ± 0.0141
Meta-learning 30.0764 ± 1.4645 0.8135 ± 0.0380 0.0040 ± 0.0017 0.8482
35% Conventional 31.5320 ± 1.5242 0.7923 ± 0.0420 0.0521 ± 0.0119
Meta-learning 32.1084 ± 1.6481 0.8553 ± 0.0379 0.0025 ± 0.0011 0.6552

Table 4. Quantitative evaluations of the reconstructions of T2 brain image associated with various
sampling ratios of radial masks. Meta-learning was trained with CS ratios of 10%, 20%, 30%, and
40% and tested with unseen ratios of 15%, 25%, and 35%. Conventional method was subjected to
regular training and testing with the same CS ratios of 15%, 25%, and 35%.

CS Ratio Methods PSNR SSIM NMSE σ (ωi )

15% Conventional 24.8921 ± 1.2356 0.6259 ± 0.0285 0.1749 ± 0.0280
Meta-learning 26.7031 ± 1.2553 0.7104 ± 0.0318 0.0205 ± 0.0052 0.9532
25% Conventional 29.0545 ± 1.1980 0.7945 ± 0.0292 0.1083 ± 0.0173
Meta-learning 30.0698 ± 0.9969 0.8164 ± 0.0235 0.0093 ± 0.0022 0.8595
35% Conventional 31.5201 ± 1.0021 0.7978 ± 0.0236 0.0815 ± 0.0129
Meta-learning 32.0683 ± 0.9204 0.8615 ± 0.0209 0.0059 ± 0.0014 0.7388

Table 5. Quantitative evaluations of the reconstructions of T2 brain image associated with various
sampling ratios of random Cartesian masks.

CS Ratio Methods PSNR SSIM NMSE σ (ωi )

10% Conventional 20.8867 ± 1.2999 0.5082 ± 0.0475 0.0796 ± 0.0242
Meta-learning 22.0434 ± 1.3555 0.6279 ± 0.0444 0.0611 ± 0.0188 0.9361
20% Conventional 22.7954 ± 1.2819 0.6057 ± 0.0412 0.0513 ± 0.0157
Meta-learning 24.7162 ± 1.3919 0.6971 ± 0.0380 0.0329 ± 0.0101 0.8320
30% Conventional 24.2170 ± 1.2396 0.6537 ± 0.0360 0.0371 ± 0.0117
Meta-learning 26.4537 ± 1.3471 0.7353 ± 0.0340 0.0221 ± 0.0068 0.6771
40% Conventional 25.3668 ± 1.3279 0.6991 ± 0.0288 0.1657 ± 0.0265
Meta-learning 27.5367 ± 1.4107 0.7726 ± 0.0297 0.0171 ± 0.0050 0.6498
 J. Imaging 2021, 7, 231 16 of 29
J. Imaging 2021, 1, 0 16 of 29

meta-learning result
conventional result
ISTA-Net+ result
meta-learning detail
conventional detail
ISTA-Net+ detail

0.02
meta-learning error

0.04

0.06

0.08

0.1

0.12

0.14

0.16

0.18

0.2
conventional error
ISTA-Net+ error
Masks

Figure 1. The pictures (from top to bottom) display the reconstruction results, zoomed-in details,
Figure 1. The pictures (from top to bottom) display the reconstruction results, zoomed-in details,
point-wise errors with a color bar, and associated radial masks for meta-learning, conventional
point-wise errors with a color bar, and associated radial masks for meta-learning, conventional
learning, and ISTA-Net+ with four different CS ratios of 10%, 20%, 30%, 40% (from left to right). The
learning, + with four different CS ratios of 10%, 20%, 30%, 40% (from left to right). The
top-right image isand ISTA-Net
the ground truth fully-sampled image.
top-right image is the ground truth fully-sampled image.
J. Imaging 2021, 7, 231 17 of 29

meta-learning result
conventional result
meta-learning detail
conventional detail

0.02
meta-learning error

0.04

0.06

0.08

0.1

0.12

0.14

0.16

0.18

0.2
conventional error
Masks

Figure 2. The pictures (from top to bottom) display the T2 brain image reconstruction results,
zoomed-in details, point-wise errors with a color bar, and associated radial masks for meta-learning
and conventional learningwith four different CS ratios of 10%, 20%, 30%, 40% (from left to right).
The top-right iage is the ground truth fully-sampled image.
J. Imaging 2021, 7, 231 18 of 29

meta-learning result
conventional result
meta-learning detail
conventional detail

0.02
meta-learning error

0.04

0.06

0.08

0.1

0.12

0.14

0.16

0.18

0.2
conventional error
Masks

Figure 3. The pictures (from top to bottom) display the T1 brain image reconstruction results,
zoomed-in details, point-wise errors with a color bar, and associated radial masks for meta-learning
and conventional learning. Meta-learning was trained with CS ratios of 10%, 20%, 30%, and 40% and
tested with three different unseen CS ratios of 15%, 25%, and 35% (from left to right). Conventional
learning was trained and tested with the same CS ratios of 15%, 25%, and 35%. The top-right image
is the ground truth fully-sampled image.
J. Imaging 2021, 7, 231 19 of 29

meta-learning result
conventional result
meta-learning detail
conventional detail

0.02
meta-learning error

0.04

0.06

0.08

0.1

0.12

0.14

0.16

0.18

0.2
conventional error
Masks

Figure 4. The pictures (from top to bottom) display the T2 brain image reconstruction results,
zoomed-in details, point-wise errors with a color bar, and associated radial masks for meta-learning
and conventional learning. Meta-learning was trained with CS ratios of 10%, 20%, 30%, and 40% and
test edwith three different unseen CS ratios of 15%, 25%, and 35% (from left to right). Conventional
learning was trained and tested with the same CS ratios of 15%, 25%, and 35%. The top-right image
is the ground truth fully-sampled image.
J. Imaging 2021, 7, 231 20 of 29

meta-learning result
conventional result
meta-learning detail
conventional detail

0
meta-learning error

0.05

0.1

0.15

0.2

0.25

0.3
conventional error
Masks

Figure 5. The pictures (from top to bottom) display the T2 brain image reconstruction results, zoomed-
in details, point-wise errors with a color bar, and associated Cartesian masks for meta-learning and
conventional learningwith four different CS ratios of 10%, 20%, 30%, and 40% (from left to right).
The top-right image is the ground truth fully-sampled image.

Next, we empirically demonstrate the convergence of Algorithm 1 in Figure 6. This

shows that the objective function value φ decreases and the PSNR value for testing data
increases steadily as the number of phases increases, which indicates that the learned
algorithm is indeed minimizing the learned function as we desired.
J. Imaging 2021, 7, 231 21 of 29

Figure 6. Convergence behavior of Algorithm 1 on T1 weighted MRI image reconstruction with four
different CS ratios using radial mask. Left: Objective function value φ versus phase number. Right:
PSNR value versus phase number.

5.5. Future Work and Open Challenges

Deep optimization-based meta-learning techniques have shown great generalizability,
but there are several open challenges that can be discussed and can potentially be addressed
in future work. A major issue is the memorization problem, since the base learner needs to
be optimized for a large number of phases and the training algorithm contains multiple
gradient steps; furthermore, the computation is very expensive in terms of time and
memory costs. In addition to reconstructing MRI through different trajectories, another
potential application for medical imaging could be multi-modality reconstruction and
synthesis. Capturing images of anatomy with multi-modality acquisitions enhances the
diagnostic information and could be cast as a multi-task problem that could benefit from
meta-learning.

6. Conclusions
In this paper, we put forward a novel deep model for MRI reconstructions via meta-
learning. The proposed method has the ability to solve multi-tasks synergistically, and
the well-trained model could generalize well to new tasks. Our baseline network is con-
structed by unfolding an LOA, which inherits the convergence property, improves the
interpretability, and promotes the parameter efficiency of the designed network structure.
The designated adaptive regularizer consists of a task-invariant learner and task-specific
meta-knowledge. Network training follows a bilevel optimization algorithm that mini-
mizes task-specific parameter ω in the upper level for the validation data and minimizes
task-invariant parameters θ in the lower level for the training data with fixed ω. The
proposed approach is the first model designed to solve the inverse problem by applying
meta-training on the adaptive regularization in the variational model. We consider the
recovery of undersampled raw data across different sampling trajectories with various
sampling patterns as different tasks. Extensive numerical experiments on various MRI
datasets demonstrate that the proposed method generalizes well at various sampling
trajectories and is capable of fast adaption to unseen trajectories and sampling patterns.
The reconstructed images achieve higher quality compared to conventional supervised
learning for both seen and unseen k-space trajectory cases.

Author Contributions: All authors are contributed equally. Conceptualization, W.B., Y.C., X.Y. and
Q.Z.; methodology, W.B., Y.C., X.Y. and Q.Z.; software, W.B. and Q.Z.; validation, Y.C. and X.Y.;
formal analysis, W.B., Y.C., X.Y. and Q.Z.; investigation, W.B., Y.C., X.Y. and Q.Z.; resources, Y.C.;
data curation, W.B. and Q.Z.; writing—original draft preparation, W.B. and Q.Z.; writing—review
and editing, X.Y. and Y.C.; visualization, X.Y. and Y.C.; supervision, Y.C.; project administration, Y.C.;
funding acquisition, X.Y. and Y.C. All authors have read and agreed to the published version of the
manuscript.
Funding: This research was partially supported by NSF grants DMS-1319050, DMS-1719932, DMS-
1818886, DMS-1925263 and University of Florida AI Catalyst Grants.
J. Imaging 2021, 7, 231 22 of 29

Informed Consent Statement: Informed consent was obtained from all subjects involved in the study.
Data Availability Statement: The data presented in this study are available on Multimodal Brain
Tumor Segmentation Challenge (BraTs) 2018 [63] at https://www.med.upenn.edu/sbia/brats2018/
data.html (accessed on 29 October 2021).
Conflicts of Interest: The authors declare no conflict of interest.

Appendix A. Convergence Analysis

We make the following assumptions regarding f and g throughout this work:
• (a1): f is differentiable and (possibly) nonconvex, and ∇ f is L f -Lipschitz continuous.
• (a2): Every component of g is differentiable and (possibly) nonconvex, and ∇g is
L g -Lipschitz continuous.
• (a3): supx∈X k∇g(x)k ≤ M for some constant M > 0.
• (a4): φ is coercive, and φ∗ = minx∈X φ(x) > −∞.
First, we state the Clark subdifferential [13] of r (x) in Lemma A1, and we show that
the gradient of rε is Lipschitz continuous in Lemma A2.

Lemma A1. Let r (x) be defined in (4); then, the Clarke subdifferential of r at x is

g j (x)
∂r (x) = { ∑ ∇gi (x)> w j + ∑ ∇gj (x)> kgj (x)k w j ∈ Rd , kΠ(w j ; C(∇gi (x)))k ≤ 1, ∀ j ∈ I0 }, (A1)
j∈ I0 j∈ I1

where I0 = { j ∈ [m] | kg j (x)k = 0}, I1 = [m] \ I0 , and Π(w; C(A)) is the projection of w onto
C(A) which stands for the column space of A.

2M2
Lemma A2. The gradient of rε is Lipschitz continuous with constant m( L g + ε ).

1
Proof. From rε (x) = ∑m 2 2 2
j=1 (kg j (x)k + ε ) − ε, it follows that

m 1
∇ r ε (x) = ∑ ∇gj (x)> gj (x)(kgj (x)k2 + ε2 )− 2 . (A2)
j =1

1
For any x1 , x2 ∈ X , we first define h(x) = g j (x)(kg j (x)k2 + ε2 )− 2 , so kh(x)k < 1.
J. Imaging 2021, 7, 231 23 of 29

kh(x1 ) − h(x2 )k (A3a)

g j (x1 ) g j (x2 )
= q −q (A3b)
kg j (x1 )k2 + ε2 kg j (x2 )k2 + ε2

g j (x1 ) g j (x1 ) g j (x1 ) g j (x2 )

= q −q +q −q (A3c)
kg j (x1 )k2 + ε2 kg j (x2 )k2 + ε2 kg j (x2 )k2 + ε2 kg j (x2 )k2 + ε2
q q 
kg j (x2 )k2 + ε2 − kg j (x1 )k2 + ε2 g (x ) − g j (x2 )
≤ g j (x1 ) q q  + qj 1 (A3d)
kg j (x1 )k2 + ε2 kg j (x2 )k2 + ε2 kg j (x2 )k2 + ε2
q q
g j (x1 ) kg j (x2 )k2 + ε2 − kg j (x1 )k2 + ε2 1
≤ q q + g j (x1 ) − g j (x2 ) (A3e)
kg j (x1 )k2 + ε2 kg j (x2 )k2 + ε2 ε
1 q q 1
≤ kg j (x2 )k2 + ε2 − kg j (x1 )k2 + ε2 + g j (x1 ) − g j (x2 ) (A3f)
ε ε
1 kg j (x2 )k2 − kg j (x1 )k2 1
≤ q q + g j (x1 ) − g j (x2 ) (A3g)
ε kg (x )k2 + ε2 + kg (x )k2 + ε2 ε
j 2 j 1

1 kg j (x2 )k + kg j (x1 )k
1
≤ q q kg j (x2 )k − kg j (x1 )k + g j (x1 ) − g j (x2 ) (A3h)
ε kg j (x2 )k2 + ε2 + kg j (x1 )k2 + ε2 ε
| {z }
<1
1 1
≤ g j (x2 ) − g j (x1 ) + g (x ) − g j (x2 ) (A3i)
ε ε j 1
2
= g (x ) − g j (x2 ) . (A3j)
ε j 1
g j ( x1 ) 1
where to obtain (A3f) we used √ < 1 and √ < 1ε .
kg j (x1 )k2 +ε2 kg j (x1 )k2 +ε2
Therefore, we have

k∇rε (x1 ) − ∇rε (x2 )k (A4a)

m
= ∑ ∇g j (x1 )> h(x1 ) − ∇g j (x2 )> h(x2 ) (A4b)
j =1
m
= ∑ ∇g j (x1 )> h(x1 ) − ∇g j (x2 )> h(x1 ) + ∇g j (x2 )> h(x1 ) − ∇g j (x2 )> h(x2 ) (A4c)
j =1
m
>
≤ ∑ ∇g j (x1 ) − ∇g j (x2 ) h(x1 ) + ∇g j (x2 )(h(x1 ) − h(x2 )) (A4d)
j =1
m
≤ ∑ ∇g j (x1 ) − ∇g j (x2 ) kh(x1 )k + k∇g j (x2 )kkh(x1 ) − h(x2 )k
j =1
m
2
≤ ∑ ∇g j (x1 ) − ∇g j (x2 ) + k∇g j (x2 )k kg j (x1 ) − g j (x2 )k by (A3) and
ε
(A4e)
j =1
2
≤ m( L g kx1 − x2 k + M · Mkx1 − x2 k), (A4f)
ε
J. Imaging 2021, 7, 231 24 of 29

where the first term of the last inequality is due to the L g -Lipschitz continuity of ∇g j . The
second term is because of kg j (x1 ) − g j (x2 )k = k∇g j (x̃)(x1 − x2 )k for some x̃ ∈ X due to
the mean value theorem and k∇g j (x̃)k ≤ supx∈X k∇g j (x)k ≤ M. Therefore, we obtain

2M2
k∇rε (x1 ) − ∇rε (x2 )k ≤ m( L g + )kx1 − x2 k. (A5)
ε

Lemma A3. Let ε, η, τt , a > 0, 0 < ρ < 1, and choose the initial x0 ∈ X . Suppose the
sequence {xt } is generated by executing Lines 3–14 of Algorithm 1 with fixed ε t = ε and that
2
0 < δ < 2/aL+ε Lε < 1 exists such that αt ≥ Lδε > 0, where Lε = L f + Lh + mL g + 2M ε and
φ∗ := minx∈X φ(x). Then, the following statements hold:
1. k∇φε (xt )k → 0 as t → ∞.
aL2ε 3
2. max{t ∈ N | k∇φε (xt+1 )k ≥ η } ≤ max{ δ2
, a }(φε (x0 ) − φ∗ + ε ).

Proof. 1. In each iteration, we compute ut+1 = zt+1 − τt σ (ωi )∇rε t (zt+1 ).

1.1. In the case the condition

1
k∇φε (xt )k ≤ akut+1 − xt k and φε (ut+1 ) − φε (xt ) ≤ − kut+1 − xt k2 (A6)
a
holds with a > 0, we put xt+1 = ut+1 , and we have φε (ut+1 ) ≤ φε (xt ).
1.2. Otherwise, we compute vt+1 = xt − αt ∇φε (xt ), where αt is found through the
line search until the criteria
1
φε (vt+1 ) − φε (xt ) ≤ − kvt+1 − xt k2 (A7)
a
holds, and then put xt+1 = vt+1 . From Lemma A2, we know that the gradient
2
∇rε (x) is Lipschitz continuous with constant m( L g + 2M
ε ). Furthermore, we
assumed in (a1) that ∇ f is L f -Lipschitz continuous. Hence, putting Lε =
2M2
L f + m( L g + ε ), we find that ∇φε is Lε -Lipschitz continuous, which implies

Lε
φε (vt+1 ) ≤ φε (xt ) + h∇φε (xt ), vt+1 − xt i + k v − xt k 2 . (A8)
2 t +1
Furthermore, by the optimality condition of

1
vt+1 = arg minh∇ f (xt ), x − xt i + σ (ωi )h∇rε (xt ), x − xt i + k x − xt k 2 ,
x 2αt

we have
1
h∇φε (xt ), vt+1 − xt i + kv − xt k2 ≤ 0. (A9)
2αt t+1
Combining (A8) and (A9) and vt+1 = xt − αt ∇φε (xt ) in line 8 of Algorithm 1
yields  
1 Lε
φε (vt+1 ) − φε (xt ) ≤ − − k vt +1 − xt k 2 . (A10)
2αt 2
Therefore, it is sufficient for αt ≤ 2/a1+ Lε for the criteria (A7) to be satisfied. This
process only take finitely many iterations since we can find a finite t such that
ρt αt ≤ 2/a1+ Lε , and through the line search, we can obtain φε (vt+1 ) ≤ φε (xt ).
Therefore, in either case of 1.1. or 1.2. where we take xt+1 = ut+1 or vt+1 , we
can obtain
φε (xt+1 ) ≤ φε (xt ), for all t ≥ 0. (A11)
J. Imaging 2021, 7, 231 25 of 29

Now, from case 1.1., (A6) gives

k∇φε (xt )k2 ≤ a2 kut+1 − xt k2 ≤ a3 (φε (xt ) − φε (ut+1 )), (A12a)

therefore if xt+1 = ut+1 we get k∇φε (xt )k2 ≤ a3 (φε (xt ) − φε (xt+1 )). (A12b)

From case 1.2. and vt+1 = xt − αt ∇φε (xt ), we have

1 1
φε (vt+1 ) − φε (xt ) ≤ − kvt+1 − xt k2 = − α2t k∇φε (xt )k2 (A13a)
a a
2 a

=⇒k∇φε (xt )k ≤ 2 φε (xt ) − φε (vt+1 ) , (A13b)
αt
a
then if xt+1 = vt+1 , we have k∇φε (xt )k2 ≤ 2 φε (xt ) − φε (xt+1 ) .


(A13c)
αt

1
Since δ
Lε ≤ αt ≤ 2/a+ Lε , we have

aL2ε
k∇φε (vt+1 )k2 ≤


2
φε (vt+1 ) − φε (xt+1 ) . (A14)
δ
aL2ε 3
Combining (A12b) and (A14) and selecting C = max{ δ2
, a }, we obtain

k∇φε (xt )k2 ≤ C φε (xt ) − φε (xt+1 ) .

(A15)

Summing up (A15) for t = 0, · · · , T, we have

T
∑ k∇φε (xt )k2 ≤ C


φε (x0 ) − φε (xT +1 ) . (A16)
t =0

Combined with the fact that φε (x) ≥ φ(x) − ε ≥ φ∗ − ε for every x ∈ X , we have

T
∑ k∇φε (xt )k2 ≤ C(φε (x0 ) − φ∗ + ε). (A17)
t =0

The right-hand side is a finite constant, and hence by letting t → ∞, we know that
k∇φε (xt )k → 0, which proves the first statement.
2. Denote κ := max{t ∈ N | k∇φε (xt+1 )k ≥ η }; then, we know that k∇φε (xt+1 )k ≥ η
for all t ≤ κ − 1. Hence, we have
κ −1 κ
κη 2 ≤ ∑ k∇φε (xt+1 )k2 = ∑ k∇φε (xt )k2 ≤ C(φε (x0 ) − φ∗ + ε). (A18)
t =0 t =1

which implies the second statement.

Lemma A4. Suppose that the sequence {xt } is generated by Algorithm 1 with an initial guess x0 .
Then, for any t ≥ 0, we have φε t+1 (xt+1 ) + ε t+1 ≤ φε t (xt ) + ε t .

Proof. To prove this statement, we can prove

φε t+1 (xt+1 ) + ε t+1 ≤ φε t (xt+1 ) + ε t ≤ φε t (xt ) + ε t . (A19)

The second inequality is immediately obtained from (A11). Now, we prove the first
inequality.
For any ε > 0, denote
q
rε,j (x) = kg j (x)k22 + ε2 − ε. (A20)
J. Imaging 2021, 7, 231 26 of 29

Since φε (x) = f (x) + σ (ωi ) ∑m

j=1 rε,j (x), it suffices to show that

rε t+1 ,j (xt+1 ) + ε t+1 ≤ rε t ,j (xt+1 ) + ε t (A21)

If ε t+1 = ε t , then the two quantities above are identical and the first inequality holds. Now,
suppose ε t+1 = γε t ≤ ε t ; then,
q q
rε t+1 ,j (xt+1 ) + ε t+1 = kg j (x)k22 + ε2t+1 ≤ kg j (x)k22 + ε2t = rε t ,j (xt+1 ) + ε t , (A22)

which implies the first inequality of (A19).

Theorem A5. Suppose that {xt } is the sequence generated by Algorithm 1 with any initial x0 ,
etol = 0 and T = ∞. Let {xtl +1 } be the subsequence that satisfies the reduction criterion in step
15 of Algorithm 1, i.e., k∇φε t (xtl +1 )k ≤ σε tl γ for t = tl and l = 1, 2, · · · . Then {xtl +1 } has at
l
least one accumulation point, and every accumulation point of {xtl +1 } is a clarke stationary point
of minx φ(x) := f (x) + σ (ωi )r (x).

Proof. By Lemma A4 and φ(x) ≤ φε (x) + ε for all ε > 0 and x ∈ X , we know that

φ(xt ) ≤ φε t (xt ) + ε t ≤ · · · ≤ φε 0 (x0 ) + ε 0 < ∞. (A23)

Since φ is coercive, we know that {xt } is bounded, and the selected subsequence {xtl +1 } is
also bounded and has at least one accumulation point.
Note that k∇φε t (xtl +1 )k ≤ σε tl γ = σε 0 γl +1 → 0 as l → ∞. Let {x p+1 } be any con-
l
vergent subsequence of {xtl +1 } and denote ε p as the corresponding ε t used in Algorithm 1
that generates x p+1 . Then, there exists x∗ ∈ X such that x p+1 → x∗ as ε p → 0, and
∇φε p (x p+1 ) → 0 as p → ∞.
Note that the Clarke subdifferential of φ at x∗ is given by ∂φ(x∗ ) = ∂ f (x∗ ) + σ (ωi )∂r (x∗ ):

g j (x̂)
∂φ(x̂) = {∇ f (x̂) + σ (ωi ) ∑ ∇gj (x̂)> wj + σ(ωi ) ∑ ∇gj (x̂)> kgj (x̂)k
j∈ I0 j∈ I1
kΠ(w j ; C(∇g j (x̂)))k ≤ 1, ∀ j ∈ I0 }, (A24)

where I0 = { j ∈ [m] | kgi (x̂)k = 0} and I1 = [m] \ I0 .

If j ∈ I0 , we have kg j (x)k = 0 ⇐⇒ g j (x) = 0: then,

g j (x) g j (x)
∂rε (x) = ∑ ∇g j (x) >   12 + ∑ ∇g j (x) >   12 (A25a)
j∈ I0 j∈ I1
kg j (x)k2 + ε2 kg j (x)k2 + ε2
g j (x)
= 0+ ∑ ∇g j (x) >   12 (A25b)
j∈ I1
kg j (x)k2 + ε2

Therefore, we obtain

g j (x)
∇φε p (x p+1 ) = ∇ f (x p+1 ) + σ(ωi ) ∑ ∇g j (x) >   12 . (A26)
j∈ I1
kg j (x)k2 + ε2p

Comparing (A24) and (A26), we can see that the first term on the right-hand side of (A26)
converges to that of (A24), due to the fact that x p+1 → x̂ and the continuity of ∇ f . Together
with the continuity of gi and ∇gi , the last term of (A24) converges to the last term of (A26)
J. Imaging 2021, 7, 231 27 of 29

as ε p → 0 and kg j (x)k > 0. Furthermore, apparently 0 is a special case of the second term
in (A26). Hence, we know that

dist(∇φε p (x p+1 ), ∂φ(x̂)) → 0,

as p → ∞. Since ∇φε p (x p+1 ) → 0 and ∂φ(x̂) is closed, we conclude that 0 ∈ ∂φ(x̂).

References
1. Li, D.; Yang, Y.; Song, Y.Z.; Hospedales, T.M. Learning to generalize: Meta-learning for domain generalization. In Proceedings of
the Thirty-Second AAAI Conference on Artificial Intelligence, New Orleans, LA, USA, 2–7 February 2018.
2. Quiñonero-Candela, J.; Sugiyama, M.; Lawrence, N.D.; Schwaighofer, A. Dataset Shift in Machine Learning; Mit Press: Cambridge,
MA, USA, 2009.
3. Ben-David, S.; Blitzer, J.; Crammer, K.; Pereira, F. Analysis of representations for domain adaptation. Adv. Neural Inf. Process.
Syst. 2007, 19, 137.
4. Balaji, Y.; Sankaranarayanan, S.; Chellappa, R. Metareg: Towards domain generalization using meta-regularization. Adv. Neural
Inf. Process. Syst. 2018, 31, 998–1008.
5. Fan, J.; Han, F.; Liu, H. Challenges of big data analysis. Natl. Sci. Rev. 2014, 1, 293–314. [CrossRef] [PubMed]
6. Day, O.; Khoshgoftaar, T.M. A survey on heterogeneous transfer learning. J. Big Data 2017, 4, 1–42. [CrossRef]
7. Munkhdalai, T.; Yu, H. Meta networks. In Proceedings of the International Conference on Machine Learning, Sydney, Australia,
6–11 August 2017; pp. 2554–2563.
8. Finn, C.; Abbeel, P.; Levine, S. Model-agnostic meta-learning for fast adaptation of deep networks. In Proceedings of the
International Conference on Machine Learning, Sydney, NSW, Australia, 6–11 August 2017; pp. 1126–1135.
9. Rusu, A.A.; Rao, D.; Sygnowski, J.; Vinyals, O.; Pascanu, R.; Osindero, S.; Hadsell, R. Meta-Learning with Latent Embedding
Optimization. In Proceedings of the International Conference on Learning Representations, Vancouver, BC, Canada, 30 April–3
May 2018.
10. Yao, H.; Huang, L.K.; Zhang, L.; Wei, Y.; Tian, L.; Zou, J.; Huang, J. Improving generalization in meta-learning via task augmenta-
tion. In Proceedings of the International Conference on Machine Learning, Virtual Event, 18–24 July 2021; pp. 11887–11897.
11. Thrun, S.; Pratt, L. Learning to learn: Introduction and overview. In Learning to Learn; Springer: Berlin/Heidelberg, Germany,
1998; pp. 3–17.
12. Hospedales, T.M.; Antoniou, A.; Micaelli, P.; Storkey, A.J. Meta-Learning in Neural Networks: A Survey. arXiv 2020,
arXiv:2004.05439.
13. Chen, Y.; Liu, H.; Ye, X.; Zhang, Q. Learnable Descent Algorithm for Nonsmooth Nonconvex Image Reconstruction. SIAM J.
Imaging Sci. 2021, 14, 1532–1564. [CrossRef]
14. Yao, H.; Wu, X.; Tao, Z.; Li, Y.; Ding, B.; Li, R.; Li, Z. Automated relational meta-learning. arXiv 2020, arXiv:2001.00745.
15. Lee, Y.; Choi, S. Gradient-based meta-learning with learned layerwise metric and subspace. In Proceedings of the International
Conference on Machine Learning, Stockholm, Sweden, 10–15 July 2018; pp. 2927–2936.
16. Huisman, M.; van Rijn, J.N.; Plaat, A. A survey of deep meta-learning. Artif. Intell. Rev. 2021, 54, 4483–4541. [CrossRef]
17. Koch, G.; Zemel, R.; Salakhutdinov, R. Siamese neural networks for one-shot image recognition. In Proceedings of the ICML
Deep Learning Workshop, Lille, France, 6– 11 July 2015; Volume 2.
18. Vinyals, O.; Blundell, C.; Lillicrap, T.; Wierstra, D. Matching networks for one shot learning. Adv. Neural Inf. Process. Syst. 2016,
29, 3630–3638.
19. Snell, J.; Swersky, K.; Zemel, R.S. Prototypical networks for few-shot learning. arXiv 2017, arXiv:1703.05175.
20. Mishra, N.; Rohaninejad, M.; Chen, X.; Abbeel, P. A simple neural attentive meta-learner. arXiv 2017, arXiv:1707.03141.
21. Ravi, S.; Larochelle, H. Optimization as a Model for Few-Shot Learning, 2016. Available online: https://openreview.net/forum?
id=rJY0-Kcll (accessed on 29 October 2021).
22. Qiao, S.; Liu, C.; Shen, W.; Yuille, A.L. Few-shot image recognition by predicting parameters from activations. In Proceedings of
the IEEE Conference on Computer Vision and Pattern Recognition, Salt Lake City, UT, USA, 18–23 June 2018; pp. 7229–7238.
23. Graves, A.; Wayne, G.; Danihelka, I. Neural turing machines. arXiv 2014, arXiv:1410.5401.
24. Rajeswaran, A.; Finn, C.; Kakade, S.; Levine, S. Meta-Learning with Implicit Gradients, 2019. Available online: https://papers.
nips.cc/paper/2019/hash/072b030ba126b2f4b2374f342be9ed44-Abstract.html (accessed on 29 October 2021)
25. Li, Z.; Zhou, F.; Chen, F.; Li, H. Meta-sgd: Learning to learn quickly for few-shot learning. arXiv 2017, arXiv:1707.09835.
26. Antoniou, A.; Edwards, H.; Storkey, A. How to train your MAML. arXiv 2018, arXiv:1810.09502.
27. Nichol, A.; Achiam, J.; Schulman, J. On first-order meta-learning algorithms. arXiv 2018, arXiv:1803.02999.
28. Finn, C.; Rajeswaran, A.; Kakade, S.; Levine, S. Online meta-learning. In Proceedings of the International Conference on Machine
Learning, Long Beach, CA, USA, 9–15 June 2019; pp. 1920–1930.
29. Grant, E.; Finn, C.; Levine, S.; Darrell, T.; Griffiths, T. Recasting gradient-based meta-learning as hierarchical bayes. arXiv 2018,
arXiv:1801.08930.
30. Finn, C.; Xu, K.; Levine, S. Probabilistic model-agnostic meta-learning. arXiv 2018, arXiv:1806.02817.
J. Imaging 2021, 7, 231 28 of 29

31. Yoon, J.; Kim, T.; Dia, O.; Kim, S.; Bengio, Y.; Ahn, S. Bayesian model-agnostic meta-learning. In Proceedings of the 32nd
International Conference on Neural Information Processing Systems, Montréal, QC, Canada, 3–8 December 2018; pp. 7343–7353.
32. Vuorio, R.; Sun, S.H.; Hu, H.; Lim, J.J. Multimodal model-agnostic meta-learning via task-aware modulation. arXiv 2019,
arXiv:1910.13616.
33. Yao, H.; Wei, Y.; Huang, J.; Li, Z. Hierarchically structured meta-learning. In Proceedings of the International Conference on
Machine Learning, Virtual, 18–24 July 2021; pp. 7045–7054.
34. Yin, M.; Tucker, G.; Zhou, M.; Levine, S.; Finn, C. Meta-Learning without Memorization. arXiv 2019, arXiv:1912.03820.
35. Jenni, S.; Favaro, P. Deep bilevel learning. In Proceedings of the European Conference on Computer Vision (ECCV), Munich,
Germany, 8–14 September 2018; pp. 618–633.
36. Chen, W.Y.; Liu, Y.C.; Kira, Z.; Wang, Y.C.; Huang, J.B. A Closer Look at Few-shot Classification. In Proceedings of the
International Conference on Learning Representations, New Orleans, LA, USA, 6–9 May 2019.
37. Li, Y.; Yang, Y.; Zhou, W.; Hospedales, T. Feature-critic networks for heterogeneous domain generalization. In Proceedings of the
International Conference on Machine Learning, Long Beach, CA, USA, 9–15 June 2019; pp. 3915–3924.
38. Rebuffi, S.A.; Bilen, H.; Vedaldi, A. Learning multiple visual domains with residual adapters. In Proceedings of the Advances in
Neural Information Processing Systems, Long Beach, CA, USA, 4–9 December 2017.
39. Triantafillou, E.; Zhu, T.; Dumoulin, V.; Lamblin, P.; Evci, U.; Xu, K.; Goroshin, R.; Gelada, C.; Swersky, K.J.; Manzagol, P.A.; et al.
Meta-Dataset: A Dataset of Datasets for Learning to Learn from Few Examples. In Proceedings of the International Conference
on Learning Representations, Addis Ababa, Ethiopia, 26–30 April 2020.
40. Yu, T.; Quillen, D.; He, Z.; Julian, R.; Hausman, K.; Finn, C.; Levine, S. Meta-world: A benchmark and evaluation for multi-
task and meta reinforcement learning. In Proceedings of the Conference on Robot Learning, Virtual, 16–18 November 2020;
pp. 1094–1100.
41. Haldar, J.P.; Hernando, D.; Liang, Z.P. Compressed-sensing MRI with random encoding. IEEE Trans. Med Imaging 2010,
30, 893–903. [CrossRef]
42. Blomgren, P.; Chan, T. Color TV: Total variation methods for restoration of vector-valued images. IEEE Trans. Image Process. 1998,
7, 304–309. [CrossRef]
43. Lundervold, A.S.; Lundervold, A. An overview of deep learning in medical imaging focusing on MRI. Z. Für Med. Phys. 2019,
29, 102–127. [CrossRef] [PubMed]
44. Liang, D.; Cheng, J.; Ke, Z.; Ying, L. Deep magnetic resonance image reconstruction: Inverse problems meet neural networks.
IEEE Signal Process. Mag. 2020, 37, 141–151. [CrossRef]
45. Sandino, C.M.; Cheng, J.Y.; Chen, F.; Mardani, M.; Pauly, J.M.; Vasanawala, S.S. Compressed sensing: From research to clinical
practice with deep neural networks: Shortening scan times for magnetic resonance imaging. IEEE Signal Process. Mag. 2020,
37, 117–127. [CrossRef]
46. McCann, M.T.; Jin, K.H.; Unser, M. Convolutional neural networks for inverse problems in imaging: A review. IEEE Signal
Process. Mag. 2017, 34, 85–95. [CrossRef]
47. Zhou, S.K.; Greenspan, H.; Davatzikos, C.; Duncan, J.S.; van Ginneken, B.; Madabhushi, A.; Prince, J.L.; Rueckert, D.; Summers,
R.M. A Review of Deep Learning in Medical Imaging: Image Traits, Technology Trends, Case Studies with Progress Highlights,
and Future Promises. 2020. [CrossRef]
48. Singha, A.; Thakur, R.S.; Patel, T. Deep Learning Applications in Medical Image Analysis. Biomed. Data Min. Inf. Retr. Methodol.
Tech. Appl. 2021, 6, 293–350.
49. Chandra, S.S.; Bran Lorenzana, M.; Liu, X.; Liu, S.; Bollmann, S.; Crozier, S. Deep learning in magnetic resonance image
reconstruction. J. Med Imaging Radiat. Oncol. 2021, 5, 564–577. [CrossRef]
50. Ahishakiye, E.; Van Gijzen, M.B.; Tumwiine, J.; Wario, R.; Obungoloch, J. A survey on deep learning in medical image
reconstruction. Intell. Med. 2021. Available online: https://www.sciencedirect.com/science/article/pii/S2667102621000061
(accessed on 29 October 2021).
51. Hammernik, K. Learning a variational network for reconstruction of accelerated MRI data. Magn. Reson. Med. 2018, 79, 3055–3071.
[CrossRef] [PubMed]
52. Aggarwal, H.K.; Mani, M.P.; Jacob, M. MoDL: Model-Based Deep Learning Architecture for Inverse Problems. IEEE Trans. Med
Imaging 2019, 38, 394–405. [CrossRef]
53. Sun, J.; Li, H.; Xu, Z. Deep ADMM-Net for compressive sensing MRI. Adv. Neural Inf. Process. Syst. 2016, 29, 10–18.
54. Zhang, J.; Ghanem, B. ISTA-Net: Interpretable optimization-inspired deep network for image compressive sensing. In Proceedings
of the IEEE Conference on Computer Vision and Pattern Recognition, Salt Lake City, UT, USA, 23 June 2018; pp. 1828–1837.
55. Lions, P.L.; Mercier, B. Splitting algorithms for the sum of two nonlinear operators. SIAM J. Numer. Anal. 1979, 16, 964–979.
[CrossRef]
56. Beck, A.; Teboulle, M. A fast iterative shrinkage-thresholding algorithm for linear inverse problems. SIAM J. Imaging Sci. 2009,
2, 183–202. [CrossRef]
57. He, K.; Zhang, X.; Ren, S.; Sun, J. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition, Las Vegas, NV, USA, 30 June 2016; pp. 770–778.
58. Mehra, A.; Hamm, J. Penalty method for inversion-free deep bilevel optimization. arXiv 2019, arXiv:1911.03432.
59. Kingma, D.P.; Ba, J. Adam: A Method for Stochastic Optimization. arXiv 2015, arXiv:1412.6980.
J. Imaging 2021, 7, 231 29 of 29

60. Glorot, X.; Bengio, Y. Understanding the difficulty of training deep feedforward neural networks. In Proceedings of the
International Conference on Artificial Intelligence and Statistics. Society for Artificial Intelligence and Statistics, Virtual Event,
13–15 April 2021.
61. Bernstein, M.A.; Fain, S.B.; Riederer, S.J. Effect of windowing and zero-filled reconstruction of MRI data on spatial resolution and
acquisition strategy. J. Magn. Reson. Imaging Off. J. Int. Soc. Magn. Reson. Med. 2001, 14, 270–280. [CrossRef]
62. Abadi, M.; Agarwal, A.; Barham, P.; Brevdo, E.; Chen, Z.; Citro, C.; Corrado, G.S.; Davis, A.; Dean, J.; Devin, M.; et al. TensorFlow:
Large-Scale Machine Learning on Heterogeneous Systems. 2015. Software. Available online: tensorflow.org (accessed on 29
October 2021).
63. Menze, B.H.; Jakab, A.; Bauer, S.; Kalpathy-Cramer, J.; Farahani, K.; Kirby, J.; Burren, Y.; Porz, N.; Slotboom, J.; Wiest, R.; et al.
The multimodal brain tumor image segmentation benchmark (BRATS). IEEE Trans. Med Imaging 2014, 34, 1993–2024. [CrossRef]
64. Abadi, M.; Barham, P.; Chen, J.; Chen, Z.; Davis, A.; Dean, J.; Devin, M.; Ghemawat, S.; Irving, G.; Isard, M.; et al. Tensorflow: A
system for large-scale machine learning. In Proceedings of the 12th {USENIX} Symposium on Operating Systems Design and
Implementation ({OSDI} 16), Savannah, GA, USA, 2–4 November 2016; pp. 265–283.
65. Glorot, X.; Bengio, Y. Understanding the difficulty of training deep feedforward neural networks. In Proceedings of the Thirteenth
International Conference on Artificial Intelligence and Statistics, Sardinia, Italy, 13–15 May 2010; pp. 249–256.
66. Wang, S.; Cheng, H.; Ying, L.; Xiao, T.; Ke, Z.; Zheng, H.; Liang, D. DeepcomplexMRI: Exploiting deep residual network for fast
parallel MR imaging with complex convolution. Magn. Reson. Imaging 2020, 68, 136–147. [CrossRef]
67. Hore, A.; Ziou, D. Image quality metrics: PSNR vs. SSIM. In Proceedings of the 2010 20th International Conference on Pattern
Recognition, Istanbul, Turkey, 23–26 August 2010; pp. 2366–2369.
68. Wang, Z.; Bovik, A.C.; Sheikh, H.R.; Simoncelli, E.P. Image quality assessment: From error visibility to structural similarity. IEEE
Trans. Image Process. 2004, 13, 600–612. [CrossRef]
69. Poli, A.; Cirillo, M. On the use of the normalized mean square error in evaluating dispersion model performance. Atmos. Environ.
Part A Gen. Top. 1993, 27, 2427–2434. [CrossRef]