Accelerating Brain MR Imaging With Multisequence and Convolutional Neural Networks

Abstract

Purpose
Magnetic resonance imaging (MRI) refers to one of the critical image modalities for diagnosis, whereas its long acquisition time limits its application. In this study, the aim was to investigate whether deep learning–based techniques are capable of using the common information in different MRI sequences to reduce the scan time of the most time‐consuming sequences while maintaining the image quality.

Method
Fully sampled T1‐FLAIR, T2‐FLAIR, and T2WI brain MRI raw data originated from 217 patients and 105 healthy subjects. The T1‐FLAIR and T2‐FLAIR sequences were subsampled using Cartesian masks based on four different acceleration factors. The fully sampled T1/T2‐FLAIR images were predicted from undersampled T1/T2‐FLAIR images and T2WI images through deep learning–based reconstruction. They were qualitatively assessed by two senior radiologists in accordance with the diagnosis decision and a four‐point scale image quality score. Furthermore, the images were quantitatively assessed based on regional signal‐to‐noise ratios (SNRs) and contrast‐to‐noise ratios (CNRs). The chi‐square test was performed, where p < 0.05 indicated a difference with statistical significance.

Results
The diagnosis decisions from two senior radiologists remained unchanged in accordance with the accelerated and fully sampled images. There were no significant differences in the regional SNRs and CNRs of most assessed regions (p > 0.05) between the accelerated and fully sampled images. Moreover, no significant difference was identified in the image quality assessed by two senior radiologists (p > 0.05).

Conclusion
Deep learning–based image reconstruction is capable of significantly expediting the brain MR imaging process and producing acceptable image quality without affecting diagnosis decisions.

Full text

Brain and Behavior
ORIGINAL ARTICLE
Accelerating Brain MR Imaging With Multisequence and
Convolutional Neural Networks
Zhanhao Mo1He Sui1Zhongwen Lv1Xiaoqian Huang2Guobin Li3Dinggang Shen2Lin Liu1
Shu Liao2
1Department of Radiology, China-Japan Union Hospital of Jilin University, Changchun, China 2Shanghai United Imaging Intelligence Co. Ltd., Shanghai,
China 3Shanghai United Imaging Co. Ltd., Shanghai, China
Correspondence: Shu Liao (shuliao2023@126.com)
Received: 12 June 2024 Revised: 5 October 2024 Accepted: 31 October 2024
Funding: This study was funded by the Development Plan for Science and Technology in Jilin Province Project (Grant/award number: YDZJ202401461ZYTS).
Keywords: accelerated imaging | convolutional neural networks | deep learning | magnetic resonance imaging
ABSTRACT
Purpose: Magnetic resonance imaging (MRI) refers to one of the critical image modalities for diagnosis, whereas its long
acquisition time limits its application. In this study, the aim was to investigate whether deep learning–based techniques are capable
of using the common information in different MRI sequences to reduce the scan time of the most time-consuming sequences while
maintaining the image quality.
Method: Fully sampled T1-FLAIR, T2-FLAIR, and T2WI brain MRI raw data originated from 217 patients and 105 healthy subjects.
The T1-FLAIR and T2-FLAIR sequences were subsampled using Cartesian masks based on four different acceleration factors.
The fully sampled T1/T2-FLAIR images were predicted from undersampled T1/T2-FLAIR images and T2WI images through
deep learning–based reconstruction. They were qualitatively assessed by two senior radiologists in accordance with the diagnosis
decision and a four-point scale image quality score. Furthermore, the images were quantitatively assessed based on regional signal-
to-noise ratios (SNRs) and contrast-to-noise ratios (CNRs). The chi-square test was performed, where p<0.05indicated a difference
with statistical significance.
Results: The diagnosis decisions from two senior radiologists remained unchanged in accordance with the accelerated and fully
sampled images. There were no significant differences in the regional SNRs and CNRs of most assessed regions (p>0.05) between
the accelerated and fully sampled images. Moreover, no significant difference was identified in the image quality assessed by two
senior radiologists (p>0.05).
Conclusion: Deep learning–based image reconstruction is capable of significantly expediting the brain MR imaging process and
producing acceptable image quality without affecting diagnosis decisions.
1 Introduction
Magnetic resonance imaging (MRI) plays a crucial role in accu-
rate disease diagnosis due to its unique advantages, such as being
radiation-free with superior soft tissue contrast and providing
complementary anatomy information through different sequence
designs. However, its slow acquisition speed leads to patient
discomfort and low hospital throughput. Therefore, there is
Abbreviations: CNR, contrast-to-noise ratio; CSF, cerebrospinal fluid; FLAIR, fluid-attenuated inversion recovery; LSGAN,least squared generative adversarial network; PACS, picture archiving and
communication system; ROI, region of interest; SNR, signal-to-noise ratio; T2WI, T2-weighted image.
This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, provided the original work is properly
cited.
© 2024 The Author(s). Brain and Behavior published by Wiley Periodicals LLC.
Brain and Behavior,2024;14:e70163
https://doi.org/10.1002/brb3.70163
1of11

an urgent need for accelerating MRI scans that have garnered
widespread attention over decades.
The most common way to expedite an MRI scan is to under-
sample the k-space from which raw MRI signals are obtained
(Twieg 1983). However, undersampling the k-space will cause
aliasing and artifacts in the resulting image since it violates
the Nyquist–Shannon constraint. Several strategies have been
proposed to address this problem. For instance, a possible way is
to explicitly enforce data redundancy (e.g., parallel imaging; Zhu
et al. 2022). Representative parallel imaging methods comprise
SENSitivity Encoding (SENSE) (Kim et al. 2022), GeneRalized
Autocalibrating Partial Parallel Acquisition (GRAPPA) (Lingala
et al. 2017), and so forth. SENSE considers the spatial sensitivities
of each coil, and the aliased images are unwrapped in the real
space. Besides, GRAPPA accelerates the reconstruction in the
complex frequency domain before the signal transforms into the
image space. Compressed sensing refers to another stream of MRI
fast reconstruction methods. It assumes that MRI data follow
the sparse representation in several transform domains, and the
image can be recovered through iterative optimization. The most
used transform domain comprises the image domain and the
wavelet domain. Furthermore, other sparsity constraints (e.g., the
dictionary-based sparsity and low-rank sparsity) were adopted.
In recent years, deep-learning techniques have made significant
strides in image analysis and reconstruction (Anwar, Abrar, and
Ullah 2023; Bin Tufail et al. 2021; Rasheed et al. 2024; Rasheed
et al. 2023; Ullah et al. 2023a, 2023b,2023c, 2024), particularly in
fast MRI reconstruction. For example, the AUtomated TransfOrm
by Manifold APproximation method, based on deep learning,
directly reconstructs images from raw signals (Ahmed et al. 2021).
The use of a deep cascade strategy has been employed for dynamic
MRI reconstruction to enforce data consistency constraints in the
frequency domain sequentially; this approach was subsequently
extended to incorporate recurrent neural networks for more
precise dynamic MRI reconstruction. In addition, the varia-
tional network has been proposed for rapid MRI reconstruction
(Wang et al. 2022). Furthermore, generative adversarial networks
(GANs) have been widely utilized for fast MRI reconstruction
(Choi et al. 2022). GAN-based frameworks rely on the evolution
of both generator and discriminator networks to automatically
enhance image contrast and appearance. Representative GAN-
based methods include de-aliasing GAN and context-aware GAN
(Joshi et al. 2022;Yangetal.2018).
Nevertheless, there remains uncertainty regarding the impact of
deep learning–based rapid MRI reconstruction within clinical
settings due to a predominant focus on publicly available data
from healthy subjects with consistent anatomical features. This
investigation evaluates both qualitatively and quantitatively how
effective deep learning–based rapid MRI reconstruction tech-
niques are for brain imaging. Our emphasis lies on expediting
fluid-attenuated inversion recovery (FLAIR) sequences, given
their significance, despite being time-consuming during routine
brain MRIs. Furthermore, leveraging multisequence information
has shown potential for enhancing deep learning–based recon-
struction quality; thus, our study employs an enhanced network
architecture. Specifically, we utilize a fully sampled T2-weighted
image (T2WI) sequence (a standard component of routine brain
MRIs acquired more swiftly than FLAIR sequences) alongside a
FIGURE 1 Flowchart of the data selection and exclusion criteria,
and the separation of the training and testing set.
downsampled FLAIR sequence as inputs for prediction by our
network.
In this study, we aim to investigate the potential of deep
learning–based techniques in harnessing shared information
across different MRI sequences from the same patient to signifi-
cantly reduce scan time for the most time-consuming sequences,
while upholding image quality for accurate diagnosis. This
study makes three primary contributions: (1) introducing a deep
learning–based reconstruction technique designed to expedite
the time-consuming FLAIR sequence using downsampled FLAIR
and fully sampled T2WI sequences, (2) evaluating the proposed
method on both the healthy subjects and the clinical patients
to assess its effectiveness and generalizability; (3) conducting
comprehensive quantitative and qualitative assessments of the
proposed method, including exploration of various acceleration
factors.
2Materials and Methods
2.1 Cohort
This study included 110 healthy adults and 220 patients (18 years
or older) who underwent brain imaging between December 8,
2018 and March 11, 2019. The raw MRI data for both groups
were stored in a workstation for over three months. Experiments
were conducted using the T2WI, T1-FLAIR, and T2-FLAIR
sequences. Five healthy subjects and three patients were excluded
due to motion artifacts and whole-body movement, respectively.
Therefore, the final analysis included 105 healthy subjects and 217
patients (Figure 1). The patient pathologies are listed in Table 1;
note that multiple pathologies may exist in a single patient.
Subsampled T1-FLAIR/T2-FLAIR images were generated using
Cartesian undersampling masks with acceleration factors of 3.0,
3.5, 4.0, and 4.5.
2.2 MRI Protocol
All patients were examined by a 3T scanner (uMR780, United
Imaging Healthcare, Shanghai, China). It includes three brain
imaging protocols using a 24-channel head coil (HNC-24). A two-
dimensional T2-weighted sequence with the following parame-
ters:TR:4220ms,TE:104.8ms,FA:125
◦,FOV:230mm×200 mm,
acquired voxel size: 0.60 mm ×0.60 mm ×5.00 mm, and
plane: axial. A two-dimensional T1-weighted FLAIR sequence
with the following parameters: TR: 2015 ms, TE: 12.6 ms,
TI: 860 ms, FA: 135◦, FOV: 230 mm ×200 mm, acquired
2of11 Brain and Behavior,2024

TABLE 1 The disease distribution of recruited patients.
Diagnoses Number
Gender Age (years,
mean ±SD)
Male Female
Encephalomalacia 9 4 5 63.33 ±8.37
Chronic cerebral
hemorrhage
96360.54±5.19
Massive Infarction 13 8 5 64.29 ±8.86
Lacunar infarction 148 73 75 62.32 ±13.61
Demyelination 95 40 55 62.49 ±15.71
Arachnoid cyst 5 2 3 60.14 ±14.95
voxel size: 0.73 mm ×0.65 mm ×5.00 mm, and plane:
axial. A two-dimensional T2-weighted FLAIR sequence with
the following parameters: TR: 8000 ms, TE: 105 ms, TI: 2465,
FA: 150 ◦,FOV:230mm×200 mm, acquired voxel size:
0.95 mm ×0.76 mm ×5.00 mm, and plane: axial.
2.3 Deep Learning–Based Reconstructions
In this study, image data were processed on a computer equipped
with Ubuntu 16.04 as the operating system, utilizing the PyTorch
platform for implementing the deep-learning reconstruction
technique, with CUDA 8.0 and cuDNN 7.0. The system was archi-
tecturally robust, featuring an Intel Xeon E5-2620 v4 processor,
clocked at 2.10 GHz, accompanied by a storage capacity of 4 TB,
128 GB of RAM. It was further empowered by 4 Nvidia TITAN XP
GPUs, each with 12 GB, to accelerate computations. The training
time was around 4 days for the proposed method.
We expanded upon the network architecture designed by Xiang
et al. (2018), implementing a fully convolutional neural network
that accepts dual inputs: the undersampled sequence targeted
for reconstruction and a fully sampled reference. Specifically,
the undersampled T1-FLAIR/T2-FLAIR sequence paired with
the fully sampled T2WI sequence was fed into the network.
Network input consisted of the real and imaginary components
of these sequences following inverse Fourier transformation, and
the goal was to output the reconstructed FLAIR image’s real and
imaginary parts. Both sequences first encountered residual blocks
with 32 channels for initial feature extraction, with consequent
concatenation to integrate reference sequence information into
the downsampling sequence’s reconstruction process. Following
concatenation, a U-net-style architecture, influenced by the work
of Ronneberger, Fischer, and Brox (2015), was employed for
multiresolution information extraction, utilizing residual blocks
together with standard convolutions. This structure incorporated
four downsampling stages with max pooling and residual blocks,
starting with 128 channels that doubled after each downsam-
pling. Each residual block consisted of a 3 ×3 convolution
layer, batch normalization, and ReLu activation. The network
training was implemented in the PyTorch DL framework, using
the ADAM optimizer with an initial rate of 0.0001, betas of
(0.9, 0.999), and the learning rate decayed by half every 100
epochs. Conversely, the upsampling path, also with four stages,
featured deconvolution and residual blocks, halving the channel
count after each upsampling. To accelerate learning, a long
skip connection facilitated the computation of residuals between
fully sampled and undersampled images. Each sequence during
scans contained multiple channels, processed sequentially by the
network. To enhance reconstruction quality, the least squared
generative adversarial network (LSGAN) training strategy was
adopted (Ren et al. 2018). Figure 2outlines these advancements.
Importantly, the framework is adaptable to scenarios with a single
available sequence, limiting inputs to just the undersampled
sequence without altering the structure except for input adjust-
ment. Nonetheless, when a fully sampled reference sequence
is accessible alongside the undersampled one (i.e., typical in
clinical settings with multiple scans), the methodology generally
achieves superior reconstruction quality, especially under high
acceleration factors, due to the valuable anatomical cues offered
by the additional fully sampled data.
Out of the 105 healthy subjects and 217 patients, 50 random
healthy subjects and 50 random patients were selected for testing.
The remaining 222 subjects (55 healthy subjects and 167 patients)
were randomly divided into a training set (consisting of 80%
of the data, with 44 healthy subjects and 133 patients) and a
validation set (20% of the data, with 11 healthy subjects and 34
patients). The training set was utilized to train the network, while
the validation set was used to optimize model parameters. Four
different acceleration factors (3.0×,3.5×, 4.0×, and 4.5×)were
considered to evaluate the proposed fast MR imaging method.
The average inference speed to reconstruct one image was 0.4 s
for our method.
2.4 Statistical Analysis and Assessment
2.4.1 Scan and Reconstruction Time Comparison
First, the scan and reconstruction time was assessed using the
conventional strategy and the proposed method. The GRAPPA
scan and reconstruction strategy is adopted in the conventional
strategy, which is one of the most used scan and reconstruction
methods. The average total time for the conventional strategy
(i.e., the scan and reconstruction time) corresponding to the
T1-FLAIR and T2-FLAIR sequences reaches 124.2 and 131.4 s,
respectively. The average total time required for the proposed
method based on the 4.5×acceleration factor reaches 38.5 s (with
a scan time of 26.2 s and a reconstruction time of 12.3 s) and 40.5 s
3of11

FIGURE 2 Network structure of the deep learning–based fast MR reconstruction framework used in this study.
(with a scan time of 27.8 s and a reconstruction time of 12.7 s),
respectively, significantly shorter than average total time required
by the conventional strategy. The reconstruction computer has an
Intel Xeon CPU E5-2620 v4 2.10-GHz processor, 4 TB of hard disk,
128 GB of RAM, and 4 Nvidia TITAN XP x12 GB GPUs.
2.4.2 Diagnosis Consistency Analysis
Two senior radiologists conducted a blinded review of fully
sampled T1-FLAIR/T2-FLAIR images and deep learning–
reconstructed images from healthy subjects or patients to make
diagnostic assessments. The consistency between the diagnostic
assessments based on the fully sampled and deep learning–
reconstructed images was compared using Cohen’s kappa
correlation coefficient to examine the interobserver agreement
in qualitative, in vivo assessment of pathologic indexes and
visibility of anatomical structures.
2.4.3 Quantitative Analysis
Regional quantitative image quality analysis was performed using
the picture archiving and communication system workstation
at our institution. Radiologists with 5 years of subspecialty
experience drew regions of interest (ROIs) using the Horos
software (https://horosproject.org/) on representative axial slices
of conventional and accelerated images with acceleration factors
of 3.0, 3.5, 4.0, and 4.5. The undersampling masks for different
acceleration factors were generated by initially including a certain
number of adjacent lowest frequency k-space lines to ensure
a fully sampled k-space region, followed by uniform random
inclusion of the remaining k-space lines based on a preset prob-
ability that achieves the desired acceleration factor on average.
The omission of k-space lines was limited to the phase encoding
direction to simulate physically achievable accelerations for 2D
sequences. ROIs were delineated for grossly normal cerebrospinal
fluid (CSF), white matter, gray matter, fat, muscle, and pathology,
as well as surrounding pathology areas in consistent anatomical
locations and sizes between conventional brain MRI and deep
learning–reconstructed sequences presented side-by-side in the
same patient-involving areas; each ROI covered more than half
the area of its corresponding lesion. Subsequently, signal-to-
noise ratios (SNRs) and contrast-to-noise ratios (CNRs) were
calculated from these ROIs. The SNRs for white matter, grey
matter, fat tissue, rectus extraocular muscle tissue, and pathologic
tissues, as well as their surrounding regions, were calculated by
dividing the average signal intensity value within each ROIplaced
on the respective tissues (SItissue) by their standard deviation
values inside each ROI (SDtissue). Standard deviation values
from within tissues were used instead of background due to
nonuniformity across regions in an accelerated sparse image.
SNR =SItissue∕SDtissue where SDtissue denotes the average signal
intensity value of ROI placed on the corresponding tissue and
SDtissue denotes the standard deviation of the corresponding tissue
inside the ROI. The CSF, white matter, gray matter, fat, rectus
extraocular muscle, pathology, and pathology surrounding CNRs
were calculated by the following equation:
CNR (tissue1, tissue2)=|
|SItissue1−SItissue2|
|
√SDtissue1
2+SDtissue2
2
where SDtissue1 and SDtissue2 denote the average signal intensity
value of ROI placed on two different tissues, while SDtissue1 and
SDtissue2 denote the standard deviation of the two tissues inside
their corresponding ROI.
Quantitative evaluation was performed by means of the chi-
square test.
4of11 Brain and Behavior,2024

2.4.4 Qualitative Analysis
The axial accelerated and fully sampled T1WI/T2WI images were
independently evaluated by two board-certified neuroradiologists
with 6 years of MR experience, specializing in neurological and
oncological imaging. The readers were blinded to all identifying
information about sequence type. Each reader independently
assessed the following six features for the respective sequence:
1. Overall image quality
2. Gray–white matter delineation
3. Lesion delineation
4. Overall diagnostic usefulness
5. Contrast fat and rectus extraocular muscle
6. Presence and severity of susceptibility artifacts
Features 1–5 were rated on a 5-point scale as follows: excellent =4,
good =3, fair =2, poor =1, and nondiagnostic =0. Feature
6 was evaluated on a 3-point scale: artifacts present and severe
enough to impair diagnostic confidence =2, artifacts present but not
impairing diagnostic confidence =1, and artifacts not present =0.
The mean scores from both raters served as the representative
value. The statistical significance of the qualitative data was
determined using the Kruskal–Wallis nonparametric test.
2.4.5 Statistical Analysis
Statistical analysis was performed using SPSS (IBM SPSS Statis-
tics 23.0, Armonk, NY). The inter-reader agreement for the
qualitative scores was evaluated using Cohen’s weighted kappa
statistics. Interobserver agreement was considered less than
chance when k<0, slight when 0.01 <k<0.20, fair when
0.21 <k<0.40, moderate when 0.41 <k<0.60, substantial
when 0.61 <k<0.80, and almost perfect when 0.81 <k<0.99.
Data regarding contrast ratio measurements are presented as
mean scores with standard deviations. The statistical signifi-
cance of contrast ratio measurements and qualitative ratings
was determined through analysis of variance and the Kruskal–
Wallis nonparametric test. pvalues less than 0.05 were deemed
statistically significant.
3Results
3.1 Diagnosis Consistency Results
For both the 50 healthy subjects and the 50 patients, both
senior radiologists arrived at the same diagnostic conclusion
based on the blinded reading experiment for the fully sampled
T1-FLAIR/T2-FLAIR data and the deep learning–reconstructed
T1/T2-FLAIR data across all acceleration factors. Consequently,
the robustness of the deep learning–based reconstruction algo-
rithm was validated. The fully sampled T1-FLAIR/T2-FLAIR
images and their deep learning–reconstructed counterparts are
depicted in Figure 3, along with the diagnostic decisions made by
the two radiologists.
3.2 Quantitative Results
The results of the quantitative analysis obtained from all the
healthy subjects and clinical patients are listed in Tables 2and 3.
Bold indicates no significant difference in data.
Based on the quantitative analysis results, it is evident that the
deep learning–reconstructed T1-FLAIR images exhibit statisti-
cally similar or even higher SNR and CNR values than the
fully sampled images for most assessment metrics (e.g., white
matter, fat, and muscle). This can be attributed to the effective
suppression of Gibbs artifacts and enhancement of anatomical
sharpness achieved through deep learning–based reconstruction.
Figure 4illustrates a representative example of T1-FLAIR image
reconstructions using zero-filled, parallel imaging, and deep
learning–based methods with an acceleration factor of 4.5×.In
addition, the fully sampled image is included for comparison
purposes. As depicted in the figure, the deep learning–based
reconstruction clearly outperforms both zero-filled and parallel
imaging reconstructions. Furthermore, it effectively suppresses
Gibbs artifacts present in the fully sampled image, particularly
in areas highlighted by green circles. This observation aligns
with quantitative assessment results presented in Tables 2and 3,
where certain metrics demonstrate superior performance of deep
learning–based reconstruction compared to fully sampled images
(e.g., SNR for white matter, fat, and muscle; CNR for white/gray
matter and white matter/CSF). The SNR for CSF was lower
after deep learning–based reconstruction compared to that of
the fully sampled T1-FLAIR image, as indicated in Table 2,
due to its weak signal strength in FLAIR sequences, which is
further distorted during the downsampling acceleration process.
However, as shown in Figure 5, despite this limitation, the
quality of the CSF region after deep learning–based reconstruc-
tion still surpasses conventional methods guiding radiologists
toward accurate diagnostic decisions. For T2-FLAIR reconstruc-
tion results, Table 2indicates that the SNR of white matter and
gray matter of the deep learning–based reconstructed results
decreased with different acceleration factors. The possible reason
for this is that T2-FLAIR was more sensitive to lesions, and
the signal strength of structural anatomies became weaker and
more vulnerable during the acceleration process. Tables 2and 3
also indicate that the differences of SNR for pathological regions
and CNR between the pathological regions and surrounding
structures compared with the fully sampled images were smaller,
partially indicating the effectiveness of the deep learning–based
reconstruction with respect to the lesion regions in T2-FLAIR
sequences. It was consistent with the results shown in the diagno-
sis consistency study in Section 3.1 that the reconstructed images
from different acceleration factors led to the same diagnosis
decision as the fully sampled images made by the two senior
radiologists.
AsshowninTables2and 3, the SNR for certain anatomical struc-
tures decreases following deep learning–based reconstruction,
falling below that of fully sampled sequences. This is primarily
attributed to the heightened susceptibility of anatomies such as
CSF, which exhibit weak signal strength in FLAIR sequences,
to distortion and artifacts resulting from the downsampling
process during acceleration. Furthermore, a qualitative analy-
sis was performed to assess interobserver agreement between
two board-certified neuroradiologists, aiming to investigate the
5of11

FIGURE 3 (a) Deep learning–based reconstruction T1-FLAIR image from 4.5×acceleration factor, (b) the corresponding full sampled T1-FLAIR
image, (c) deep learning–based reconstruction T2-FLAIR image with 4.5×acceleration factor, and (d) the corresponding full sampled T2-FLAIR image.
Both radiologists made the same diagnosis decision to the deep learning–reconstructed and full sampled T1-FLAIR image as normal subject, and the
deep learning–reconstructed and full sampled T2-FLAIR image has bilateral basal ganglia.
TABLE 2 SNR comparison between the fully sampled T1/T2-FLAIR and deep learning–based reconstruction results with respect to different
acceleration factors for the volunteer and clinical group.
SNR Full DL 3.0×DL 3.5×DL 4.0×DL 4.5×p
T1-FLAIR volunteer
White matter 24.73 ±4.7 31.72 ±5.51 36.81 ±7.56 34.56 ±7.55 30.92 ±7.33 <0.01
Gray matter 12.91 ±3.68 12.93 ±3.97 13.82 ±5.10 13.28 ±4.12 11.37 ±3.00 0.510
CSF 6.32 ±0.84 3.68 ±0.69 3.55 ±0.70 3.59 ±0.64 3.31 ±0.74 <0.01
Fat 12.08 ±2.22 14.91 ±3.55 15.42 ±4.37 17.22 ±4.70 16.01 ±4.36 <0.05
Muscle 8.08 ±3.30 9.39 ±3.93 9.91 ±4.21 10.01 ±4.06 10.71 ±4.71 <0.01
T1-FLAIR clinical
White matter 28.04 ±4.94 34.63 ±8.03 38.80 ±7.58 36.30 ±7.25 31.32 ±6.25 <0.01
Gray matter 14.88 ±3.58 13.49 ±3.62 16.08 ±7.29 14.48 ±4.79 13.45 ±3.74 <0.05
CSF 7.34 ±0.77 4.28 ±0.90 4.32 ±0.76 4.03 ±0.72 3.89 ±0.64 <0.01
Fat 14.76 ±3.88 16.63 ±4.49 20.81 ±6.27 18.20 ±5.28 17.77 ±4.44 <0.05
Muscle 16.35 ±6.72 19.52 ±9.07 22.09 ±10.46 22.29 ±10.6 21.00 ±9.15 <0.01
Pathology 15.37 ±11.77 15.74 ±13.20 17.93 ±13.32 17.05 ±14.2 16.05 ±10.5 0.854
Pathology surrounding 25.02 ±9.84 26.03 ±7.95 35.26 ±12.44 30.60 ±14.27 26.97 ±8.45 <0.01
T2-FLAIR volunteer
White matter 22.08 ±4.01 22.83 ±5.14 22.12 ±4.48 23.05 ±4.44 22.84 ±4.73 0.745
Gray matter 23.37 ±6.54 23.59 ±6.02 22.77 ±6.27 22.90 ±5.57 21.91 ±5.02 0.679
CSF 5.75 ±1.55 4.77 ±0.89 4.70 ±0.96 4.74 ±0.91 4.64 ±1.02 <0.01
Fat 9.38 ±2.14 9.62 ±2.47 9.76 ±2.40 9.46 ±2.35 8.93 ±2.33 0.884
Muscle 5.48 ±1.46 5.41 ±1.48 5.46 ±1.75 5.50 ±1.93 5.73 ±1.32 0.466
T2-FLAIR clinical
White matter 19.55 ±2.57 16.89 ±1.97 17.90 ±2.41 17.69 ±2.64 16.03 ±2.12 <0.01
Gray matter 20.53 ±3.06 17.68 ±3.15 19.07 ±3.35 18.39 ±3.74 16.13 ±3.05 <0.01
CSF 6.18 ±1.12 5.60 ±0.73 5.56 ±0.71 5.43 ±0.70 5.17 ±0.62 <0.01
Fat 7.61 ±1.22 7.83 ±1.57 8.04 ±1.47 7.59 ±1.25 7.32 ±1.30 0.637
Muscle 7.36 ±1.79 7.60 ±1.77 7.55 ±1.90 7.23 ±1.70 7.12 ±1.65 0.104
Pathology 11.89 ±6.78 11.19 ±5.23 11.36 ±5.34 11.44 ±4.71 11.02 ±4.97 0.950
Pathology surrounding 19.26 ±3.85 16.97 ±3.13 18.46 ±3.57 17.82 ±3.35 15.56 ±2.53 <0.01
Note: Bold indicates no significant difference data.
Abbreviations: 3.0×=downsampling factor of 3.0, 3.5×=downsampling factor of 3.5, 4.0×=downsampling factor of 4.0, 4.5×=downsampling factor of 4.5,
CSF =cerebrospinal fluid, DL =deep learning–based reconstruction.
6of11 Brain and Behavior,2024

TABLE 3 CNR comparison between the fully sampled T1/T2-FLAIR and deep learning–based reconstruction results with respect to different
acceleration factors for the volunteer and clinical group.
CNR Full DL 3.0×DL 3.5×DL 4.0×DL 4.5×p
T1-FLAIR volunteer
White/gray matter 9.05 ±2.20 10.36 ±2.64 11.62 ±3.66 10.77 ±2.78 9.55 ±2.10 <0.01
White matter/CSF 18.43 ±3.95 22.55 ±3.87 25.75 ±5.68 24.97 ±5.33 21.74 ±5.65 <0.01
Gray matter/CSF 7.74 ±2.4 8.55 ±2.30 9.16 ±2.87 9.15 ±2.58 7.63 ±1.92 <0.01
Fat/muscle 4.23 ±2.23 4.99 ±2.68 5.38 ±2.95 5.34 ±2.84 5.65 ±3.29 0.117
T1-FLAIR clinical
White/gray matter 7.99 ±1.50 8.47 ±1.99 9.63 ±3.31 8.89 ±2.40 8.54 ±3.27 <0.05
White matter/CSF 20.34 ±3.37 23.65 ±4.76 27.22 ±5.08 25.23 ±5.11 22.52 ±3.54 <0.01
CSF/pathology 9.11 ±5.57 9.73 ±5.47 12.02 ±7.3 11.16 ±7.23 10.57 ±5.99 <0.05
Pathology/surrounding 8.16 ±6.57 8.44 ±8.77 8.23 ±6.85 7.74 ±6.91 7.53 ±6.98 <0.05
Gray matter/CSF 9.37 ±2.09 9.39 ±2.38 10.93 ±3.69 10.11 ±2.64 9.44 ±2.41 0.174
Fat/muscle 9.87 ±4.37 11.95 ±5.89 13.25 ±6.47 13.23 ±6.40 12.58 ±5.70 0.971
T2-FLAIR volunteer
White/gray matter 1.74 ±1.15 1.82 ±1.28 1.75 ±1.30 1.76 ±1.20 1.65 ±1.17 0.974
White matter/CSF 5.72 ±3.99 5.96 ±1.38 5.72 ±1.42 5.93 ±1.34 5.67 ±1.30 0.944
Gray matter/CSF 6.33 ±4.89 6.54 ±1.38 6.25 ±1.51 6.50 ±1.44 6.19 ±1.38 0.952
Fat/muscle 2.71 ±1.12 2.95 ±1.11 2.97 ±1.19 2.91 ±1.04 2.87 ±0.97 0.772
T2-FLAIR clinical
White/gray matter 1.94 ±1.05 1.73 ±0.93 1.87 ±1.01 1.85 ±1.03 1.61 ±0.91 0.474
White matter/CSF 12.03 ±1.98 10.63 ±1.40 11.40 ±1.56 11.07 ±1.61 10.02 ±1.40 <0.01
CSF/pathology 9.19 ±4.76 8.59 ±3.71 8.91 ±3.98 8.91 ±3.61 8.46 ±3.64 <0.01
Pathology/surrounding 3.36 ±2.82 2.93 ±2.16 3.01 ±2.24 2.97 ±2.05 2.83 ±1.97 0.999
Gray matter/CSF 13.46 ±1.86 11.81 ±1.64 12.83 ±1.76 12.30 ±1.98 10.80 ±1.55 0.905
Fat/muscle 1.18 ±1.06 1.18 ±1.08 1.21 ±1.17 1.19 ±1.06 1.15 ±1.14 0.811
Note: Bold indicates no significant difference data.
Abbreviations: 3.0×=downsampling factor of 3.0, 3.5×=downsampling factor of 3.5, 4.0×=downsampling factor of 4.0, 4.5×=downsampling factor of 4.5,
DL =deep learning–based reconstruction.
FIGURE 4 Typical example of (a) zero-filled reconstruction, (b) parallel imaging reconstruction,and (c) deep learning–based reconstruction result
of T1-FLAIR image with 4.5×acceleration factor. (d) This shows the full sampled reconstruction result. Note that Gibbs artifacts are suppressed by the
deep learning–based reconstruction result compared to the full sampled image, and significant differences are highlighted by the green circle.
7of11

FIGURE 5 (a) Deep learning–based reconstruction of T2-FLAIR images with a 4.5×acceleration factor; (b) VARNET reconstructed T2-FLAIR
image, showing areas of apparent signal enhancement in the cerebral cortex (thin arrows), while artifact correction remains suboptimal (thick
arrows); (c) CycleGAN-based network reconstructed T2-FLAIR image, exhibiting distortion in both signal and extent of the lesion; (d) deep learning–
based reconstruction of T1-FLAIR images with a 4.5×acceleration factor; (e) VARNET reconstructed T1-FLAIR image; (f ) CycleGAN-based network
reconstructed T1-FLAIR image, the artifact correction in the last two images is suboptimal, and their overall sharpness remains inadequate.
potential loss of image signal impacting diagnostic decisions. The
qualitative findings are detailed in the subsequent section.
3.3 Qualitative Results
Table 4lists the results of the subjective image quality analysis
obtained with interobserver agreement from all the healthy
subjects and patients. The diagnostic interobserver agreement
was all above the moderate level (i.e., k>0.41) between the
conventional and deep learning–based reconstruction results of
different acceleration factors. In this study, overall image quality,
motion artifacts, other artifacts, and image pulsation artifacts
were assessed.
From the qualitative assessment results, it is observed that both
the T1-FLAIR and T2-FLAIR deep learning–reconstructed images
from different acceleration factors have statistically very similar
scores as the fully sampled images of all the assessment metrics. It
reflects that the reconstructed images from different acceleration
factors have very similar image quality as the fully sampled
images for diagnosis purposes with respect to radiologists, includ-
ing the pathological regions for the clinical patients, as listed in
Table 4. This result also matches with the diagnosis consistency
study that both senior radiologists made the same diagnosis
decision with respect to all the subjects (including both the
healthy subjects and the clinical patients) of different acceleration
factors. The accelerated reconstruction images produced by our
proposed method demonstrated significantly superior subjective
image quality scores compared to those generated by VARNET
and the CycleGAN-based network (Oh, Lee, and Ye 2021;Sriram
et al. 2020), as shown in Table 5and Figure 5.
4 Discussion
The acceleration of MR scans has emerged as a pivotal focus
within medical imaging due to its potential for significantly
enhancing clinical workflows and patient satisfaction. This
investigation leveraged a deep learning–based approach along-
side multisequence information for expediting routine brain
T1-FLAIR/T2-FLAIR sequence scans while systematically eval-
uating reconstructed images across varying acceleration factors.
The primary advantages of deep learning are evident: It operates
as a purely data-driven method capable of autonomously deter-
mining optimal image recovery pathways for diverse acceleration
factors and sequences without reliance on explicit prior assump-
tions inherent in traditional acceleration techniques (e.g., signal
sparsity assumptions utilized in compressed sensing). Rigorous
evaluations were conducted to appraise image quality result-
ing from deep learning–based reconstruction methods. Initial
scrutiny focused on discerning whether these reconstructions led
to differential diagnostic decisions compared with fully sampled
radiologist-made images. Furthermore, quantitative assessments
employed SNR and CNR measurements across distinct anatom-
ical regions for comparing deep learning–reconstructed ver-
sus fully sampled imagery. Subsequently, qualitative evaluation
entailed comprehensive scoring of image quality across multiple
dimensions by expert radiologists.
Theproposedmethodoffersseveraladvantagesoverotherstate-
of-the-art deep learning–based methods. First, instead of solely
accelerating a single sequence as many other fast reconstruction
methods, the proposed approach incorporates an additional fully
sampled sequence (T2WI in this study) that can be acquired
relatively quickly to assist in reconstructing time-consuming
8of11 Brain and Behavior,2024

TABLE 4 Subjective image quality evaluation and comparison between the conventional and deep learning–based reconstruction results with
respect to different acceleration factors.
Full DL 3.0×DL 3.5×DL 4.0×DL 4.5×P
T1-FLAIR clinical
Overall image quality 3.42 ±0.50 3.34 ±0.48 3.52 ±0.51 3.58 ±0.50 3.44 ±0.50 0.14
Gray–white matter
delineation
3.40 ±0.50 3.36 ±0.49 3.50 ±0.51 3.46 ±0.50 3.46 ±0.50 0.65
Lesion delineation 3.52 ±0.51 3.38 ±0.49 3.44 ±0.50 3.46 ±0.50 3.42 ±0.50 0.71
Overall diagnostic usefulness 3.44 ±0.50 3.38 ±0.49 3.38 ±0.49 3.48 ±0.51 3.34 ±0.48 0.64
Contrast fat and rectus
extraocular muscle
3.46 ±0.50 3.40 ±0.50 3.38 ±-0.49 3.50 ±0.51 3.52 ±0.51 0.56
Presence and severity of
susceptibility artifacts
0.06 ±0.24 0.04 ±0.20 0.04 ±0.20 0.04 ±0.20 0.04 ±0.20 0.99
T2-FLAIR clinical
Overall image quality 3.54 ±0.50 3.54 ±0.50 3.58 ±0.50 3.50 ±0.51 3.56 ±0.50 0.95
Gray–white matter
delineation
3.54 ±0.50 3.56 ±0.50 3.56 ±0.50 3.50 ±0.51 3.50 ±0.51 0.95
Lesion delineation 3.52 ±0.51 3.50 ±0.51 3.54 ±0.50 3.50 ±0.51 3.48 ±0.51 0.98
Overall diagnostic usefulness 3.48 ±0.51 3.48 ±0.51 3.54 ±0.50 3.50 ±0.51 3.50 ±0.51 0.98
Contrast fat and rectus
extraocular muscle
3.52 ±0.51 3.50 ±0.51 3.46 ±0.50 3.52 ±0.51 3.54 ±0.50 0.95
Presence and severity of
susceptibility artifacts
0.04 ±0.20 0.06 ±0.24 0.06 ±0.24 0.04 ±0.20 0.04 ±0.20 0.97
T1-FLAIR volunteer
Overall image quality 3.50 ±0.51 3.50 ±0.51 3.56 ±0.50 3.48 ±0.51 3.54 ±0.50 0.93
Gray–white matter
delineation
3.54 ±0.50 3.50 ±0.51 3.50 ±0.51 3.48 ±0.51 3.54 ±0.50 0.97
Overall diagnostic usefulness 3.52 ±0.51 3.50 ±0.51 3.52 ±0.51 3.46 ±0.50 3.52 ±0.51 0.97
Contrast fat and rectus
extraocular muscle
3.54 ±0.50 3.48 ±0.51 3.52 ±0.51 3.48 ±0.51 3.52 ±0.51 0.97
Presence and severity of
susceptibility artifacts
0.04 ±0.20 0.06 ±0.24 0.04 ±0.20 0.06 ±0.24 0.06 ±0.24 0.98
T2-FLAIR volunteer
Overall image quality 3.58 ±0.50 3.52 ±0.51 3.48 ±0.51 3.46 ±0.50 3.52 ±0.51 0.80
Gray–white matter
delineation
3.50 ±0.501 3.52 ±0.51 3.48 ±0.51 3.46 ±0.50 3.52 ±0.51 0.97
Overall diagnostic usefulness 3.50 ±0.51 3.50 ±0.51 3.50 ±0.51 3.48 ±0.51 3.48 ±0.51 1.00
Contrast fat and rectus
extraocular muscle
3.40 ±0.50 3.52 ±0.51 3.50 ±0.51 3.50 ±0.51 3.46 ±0.50 0.73
Presence and severity of
susceptibility artifacts
0.06 ±0.24 0.04 ±0.20 0.04 ±0.20 0.04 ±0.20 0.04 ±0.20 0.99
Abbreviations: 3.0×=downsampling factor of 3.0, 3.5×=downsampling factor of 3.5, 4.0×=downsampling factor of 4.0, 4.5×=downsampling factor of 4.5,
DL =deep learning–based reconstruction.
TABLE 5 Qualitative assessment of different reconstruction methods under the 4.5×acceleration factor.
Comparison
Network 1
Comparison
Network 2
Proposed method
(DL 4.5×)p
T1-FLAIR 3.21 ±0.14 3.36 ±0.17 3.67 ±0.24 <0.01
T2-FLAIR 3.22 ±0.11 3.32 ±0.14 3.66 ±0.12 <0.01
9of11

sequences (T1-FLAIR and T2-FLAIR in this study). The rich
structural and anatomical details present in the fully sampled
sequence significantly enhance the recovery of downsampled
sequences, particularly at high acceleration factors. Second, the
use of residual blocks in the reconstruction network ensures
flexible information transfer between different network lay-
ers and reduces convergence challenges, especially for very
deep networks. In addition, we employed a LSGAN training
strategy, which demonstrates greater robustness compared to
conventional cross-entropy loss training strategies used by other
GAN-based deep learning–reconstruction methods (Mao et al.
2017).
Our model addressed challenging factors such as different clinical
settings and patient demographic differences mainly in two ways.
First, the patient cohort used to train the reconstruction model
was carefully selected to cover as many clinical settings as possi-
ble, as both the healthy subjects and the patients who underwent
brain scans were included between December 8, 2018 and March
11, 2019, where no specific disease was discarded or considered in
a specific manner. The healthy subjects and the patients included
in this study were also from different demographics admitted
to the hospital, and no specific selection was performed. In this
way, the model could be adapted to different clinical settings and
patient demographics. Second, the reconstruction method was
carefully designed to retain as many anatomical details as possible
from the multiple sequences, especially from the fully sampled
reference sequence, which can provide helpful information to the
network to reconstruct the downsampled sequence. This design
was not limited to specific scanner and can be generalized to
different vendors and machines with different field strengths
(e.g., 1.5 T), though in this study, data from the 3T scanner of
United Imaging Healthcare were used
One important point to consider is how the method can
be adapted to larger datasets, such as more anatomies and
sequences. Techniques such as adjusting the size of the mini-
batch, data shuffling, and collaborative learning can be used to
handle larger datasets (Anwar, Abrar, and Ullah 2023; Ullah et al.
2023a, 2023b,2023c, 2024), and it will not fundamentally alter the
time complexity of our approach. Also, to accommodate different
image qualities, more data augmentation techniques could be
applied to the original training images to increase the abundance
of training materials and boost the generalizability of the model.
To integrate our method with diverse MRI hardware and software
infrastructures, we also need to overcome several potential
roadblocks, such as differences in MRI scanner manufacturer
specifications, software interoperability, and data formatting
issues. One potential strategy is to develop intelligent software
wrappers that can intelligently adjust the model’s parameters
based on the individual MRI system’s characteristics. It can
significantly enhance our method’s flexibility, enabling it to
account for variations in image quality and noise levels inherent
to different scanner setups. Furthermore, close collaboration
with MRI scanner manufacturers could facilitate the inclusion
of our method within their software suites, reducing the need
for separate integration steps. Such collaborations might involve
sharing specifications, joint testing, and developing dedicated
plugin systems that marry the manufacturer’s MRI software with
our AI model.
This study has several limitations. First, the results were obtained
using a 3T MRI scanner, and it remains uncertain whether the
deep learning–based acceleration reconstruction method would
yield similar performance on a 1.5 T MRI as demonstrated in
this study. Second, while the investigation focused on the brain
region using T1-FLAIR and T2-FLAIR sequences, further com-
prehensive assessments are required to explore the performance
of deep learning–based acceleration reconstruction methods
across other body regions and sequences. It should be noted
that the computational complexity inherent in our AI-based
MRI reconstruction method could involve higher computational
overhead compared to traditional techniques. To mitigate this,
further optimization of our AI-based algorithm and dedicated
hardware acceleration, such as GPUs, was required to strike
a balance between computational efficiency and scalability.
In clinical applications, it is important to make sure critical
diagnosis information was not lost in the reconstructed images,
which is the basic principle in the trade-offs between scanning
speed and image quality. The proposed method might fail for
patients with unique pathologies, which are not well-represented
in the training sets. The dataset used to train and evaluate
the reconstruction model was carefully selected to cover as
many clinical settings as possible, as both the healthy subjects
and the patients who underwent brain scans were included
between December 8, 2018 and March 11, 2019, where no specific
disease was discarded or considered in a specific manner. The
healthy subjects and the patients included in this study were also
naturally admitted from different demographics to the hospital
during that period, and no specific selection was performed.
However, the generalizability and robustness of our method with
respect to more geographic and demographic variations are yet to
be verified.
Future research directions include extending the proposed
method to other commonly scanned anatomies such as spines,
knees, shoulders, and abdomen. Moreover, we can extend the
proposed method to more MRI sequences, such as diffusion-
weighted imaging and dynamic contrast-enhanced sequences.
Other possible network structure designs, such as stable diffu-
sion, could also be explored. It will also be important to investigate
how to incorporate the Internet of Things (IoT) into AI-driven
MRI image reconstruction. It not only broadens the scope of
enhanced imaging applications but also paves the way for a
more interconnected healthcare system. IoT enables the seamless
acquisition and transmission of patient data from MRI machines
to AI algorithms in real-time, facilitating faster reconstruction
and even remote analysis and evaluations. This integration
allows for intelligent scheduling for reconstruction resources,
maintaining privacy and security, and driving innovation in
medical imaging technology. Also, more sites may be covered and
more patients with different diseases may be recruited to evaluate
the proposed method. Finally, MRI machines, even when oper-
ating at the same field strength, can display differences due to
variations in hardware components and vendor-specific software
implementations. This heterogeneity poses a challenge for AI
models trained on data from a particular scanner, potentially
leading to inconsistencies in findings across different imaging
centers. To counter this, transfer learning and more generalized
AI algorithms are crucial for maintaining diagnostic accuracy
regardless of imaging hardware, which is also one of the further
research directions.
10 of 11 Brain and Behavior,2024

Author Contributions
Zhanhao Mo: conceptualization, data curation, formal analysis. He
Sui: data curation, visualization, writing–original draft. Zhongwen Lv:
resources, software, supervision. Xiaoqian Huang: conceptualization,
investigation, methodology, project administration. Guobin Li: investiga-
tion, methodology, validation. Dinggang Shen: supervision, validation.
Lin Liu: project administration, resources, supervision, visualization.
Shu Liao: conceptualization, investigation, methodology, project admin-
istration, supervision, visualization, writing–review and editing.
Ethics Statement
This retrospective study gained approval from the local institutional
review board (IRB), with the IRB number of 20201112.
Data Availability Statement
The data and code for this project are confidential but may be obtained
from the corresponding author upon reasonable request.
Peer Review
The peer review history for this article is available at https://publons.com/
publon/10.1002/brb3.70163.
References
Ahmed, B., F. Rahman, N. Hooten, F. Farahmandi, and M. Tehranipoor.
2021. “AutoMap: Automated Mapping of Security Properties Between
Different Levels of Abstraction in Design Flow.” Paper presented at
40th IEEE/ACM International Conference on Computer Aided Design
(ICCAD), Jeju Island, Republic of Korea.
Anwar, R. W., M. Abrar, and F. Ullah. 2023. “Transfer Learning in Brain
Tumor Classification: Challenges, Opportunities, and Future Prospects.”
Paper presented at 2023 14th International Conference on Informa-
tion and Communication Technology Convergence (ICTC), Jeju Island,
Republic of Korea.
Bin Tufail, A., I. Ullah, W. U. Khan, et al. 2021. “Diagnosis of Diabetic
Retinopathy Through Retinal Fundus Images and 3D Convolutional
Neural Networks With Limited Number of Samples.” Wireless Commu-
nications & Mobile Computing 2021: 6013448.
Choi, S.-H. J.-M. Shin P. Liu, and Y.-H. Choi. 2022. “ARGAN: Adversari-
ally Robust Generative Adversarial Networks for Deep Neural Networks
Against Adversarial Examples.” IEEE Access 10: 33602–33615.
Joshi, I. T. Prakash B. S. Jaiswal, et al. 2022. “Context-Aware Restoration
of Noisy Fingerprints.” IEEE Sensors Letters 6, no. 10: 6003704.
Kim, J. N. Seo H. Bae, et al. 2022. “Comparison of Sensitivity Encoding
(SENSE) and Compressed Sensing-SENSE for Contrast-Enhanced T1-
Weighted Imaging in Patients With Crohn Disease Undergoing MR
Enterography.” American Journal of Roentgenology 218, no. 4: 678–686.
Lingala, S. G. Y. Zhu Y.Lim, et al. 2017. “Feasibility of Through-Time Spiral
Generalized Autocalibrating Partial Parallel Acquisition for Low Latency
Accelerated Real-Time MRI of Speech.” Magnetic Resonance in Medicine
78, no. 6: 2275–2282.
Mao, X. Q. Li H. Xie, R. Y. K. Lau, Z. Wang, and S. P. Smolley. 2017.
“Least Squares Generative Adversarial Networks.” Paperp resentedat 16th
IEEE International Conference on Computer Vision (ICCV), Venice,Italy,
October 22–29, 2017.
Oh, G. J. E. Lee, and J. C. Ye. 2021. “Unpaired MR Motion Artifact
Deep Learning Using Outlier-Rejecting Bootstrap Aggregation.” IEEE
Transactions on Medical Imaging 40, no. 11: 3125–3139.
Rasheed, Z. Y.-K. Ma I. Ullah, et al. 2023. “Brain Tumor Classifica-
tion From MRI Using Image Enhancement and Convolutional Neural
Network Techniques.” Brain Sciences 13, no. 9: 1320.
Rasheed, Z., Y.-K. Ma I. Ullah M. Al-Khasawneh S. S. Almutairi,
and M. Abohashrh. 2024. “Integrating Convolutional Neural Networks
With Attention Mechanisms for Magnetic Resonance Imaging-Based
Classification of Brain Tumors.” Bioengineering 11, no. 7: 701.
Ren, K. L. Meng C. Fan, and P. Wang. 2018. “Least Squares DCGAN Based
Semantic Image Inpainting.” Paper presented at 5th IEEE International
Conference on Cloud Computing and Intelligence Systems (CCIS),
Nanjing, People’s Republic of China , November 23–25, 2018.
Ronneberger, O. P. Fischer, and T. Brox. 2015. “U-Net: Convolutional
Networks for Biomedical Image Segmentation.” Paper presented at 18th
International Conference on Medical Image Computing and Computer-
Assisted Intervention (MICCAI), Munich, Germany, October 5–9, 2015.
Sriram, A. J. Zbontar T. Murrell, et al. 2020. “End-to-End Variational
Networks for Accelerated MRI Reconstruction.” In Medical Image Com-
puting and Computer Assisted Intervention—MICCAI 2020, edited by A.
L. Martel, P. Abolmaesumi, D. Stoyanov, et al. 64–73. Cham: Springer.
Twieg, D. B. 1983. “The k-Trajectory Formulation of the NMR Imag-
ing Process With Applications in Analysis and Synthesis of Imaging
Methods.” Medical Physics 10, no. 5: 610–621.
Ullah, F. M. Nadeem M. Abrar, et al. 2023a. “Brain Tumor Segmentation
From MRI Images Using Handcrafted Convolutional Neural Network.”
Diagnostics 13, no. 16: 2650.
Ullah, F. M. Nadeem M. Abrar,et al. 2023b. “Evolutionary Model for Brain
Cancer-Grading and Classification.” IEEE Access 11: 126182–126194.
Ullah, F. M. Nadeem, and M. Abrar. 2024. “Revolutionizing Brain Tumor
Segmentation in MRI With Dynamic Fusion of Handcrafted Features and
Global Pathway-Based Deep Learning.” KSII Transactions on Internet and
Information Systems 18, no. 1: 105–125.
Ullah, F. M. Nadeem M. Abrar F. Amin A. Salam, and S. Khan. 2023c.
“Enhancing Brain Tumor Segmentation Accuracy Through Scalable
Federated Learning With Advanced Data Privacyand Security Measures.”
Mathematics 11, no. 19: 4189.
Wang, S. Y. Yang J. Sun, and Z. Xu. 2022. “Variational HyperAdam: A
Meta-Learning Approach to Network Training.” IEEE Transactions on
Pattern Analysis and Machine Intelligence 44, no. 8: 4469–4484.
Xiang, L. Y. Chen, W. Chang, et al. 2018. “Ultra-Fast T2-Weighted MR
Reconstruction Using Complementary T1-Weighted Information.” In
Medical Image Computing and Computer Assisted Intervention—MICCAI
2018, edited by A. Frangi, J. Schnabel, C. Davatzikos, C. Alberola-López,
and G. Fichtinger, 215–223. Cham: Springer.
Yang, G., S. Yu H. Dong, et al. 2018. “DAGAN: Deep De-Aliasing
Generative Adversarial Networks for Fast Compressed Sensing MRI
Reconstruction.” IEEE Transactions on Medical Imaging 37, no. 6: 1310–
1321.
Zhu, W., B. Guan S. Wang M. Zhang, and Q. Liu. 2022. “Universal
Generative Modeling for Calibration-Free Parallel MR Imaging.” Paper
presented at 19th IEEE International Symposium on Biomedical Imaging
(IEEE ISBI), Kolkata, India, March 28–31, 2022.
11 of 11