MRI Super-Resolution using Implicit Neural Representation with
Frequency Domain Enhancement
SHUANGMING MAO
SEIICHIRO KAMATA
shuangming@suou.waseda.jp
kamata@kamata-lab.org
Waseda University
Fukuoka, Japan
ABSTRACT
High resolution (HR) Magnetic Resonance Imaging (MRI) is a popu-
lar diagnostic tool, which provides detail structural information and
rich textures, beneting accurate diagnosis and disease detection.
However, obtaining HR MRI remains a challenge due to longer
scan time and lower peak signal-to-noise ratio (PSNR). Recently,
Single Image Super-Resolution (SISR) has generated interest, which
shows promising ability for recovering an HR image only relies on
a Low Resolution (LR) image. MR images have some characteristics
dierent with natural images: derived from frequency domain, sim-
pler textures and structural information. However, Most of previous
methods treat MR images as same as natural images, they only ap-
ply SR methods on natural images to MR images and fail to preserve
low-frequency information and capture high-frequency details. In
this paper, we mimic the process of an MRI machine produces an
MRI in practice and propose an Implicit Neural Representation
based module, which enable reconstruct high frequency contents
eectively while preserving low frequency contents unchanged.
Moreover, vanilla L1 loss cannot reect the dierences for each
frequency, to address this problem, we design a frequency loss to dis-
entangle each frequency and calculate the dierences respectively.
Finally, to further capture high frequency contents, we propose
High-Frequency Pixel Loss, which can decouples the HF contents
from pixel domain and emphasize the HF dierences between SR
and HR images. Extensive experiments show the eectiveness of
our proposed method in terms of visual quality and PSNR score,
which produces sharper edges and clearer details compared to pre-
vious works.
CCS CONCEPTS
• Computing methodologies → Biometrics.
KEYWORDS
MRI, Super-Resolution, Implicit Neural Representation
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for prot or commercial advantage and that copies bear this notice and the full citation
on the rst page. Copyrights for components of this work owned by others than ACM
must be honored. Abstracting with credit is permitted. To copy otherwise, or republish,
to post on servers or to redistribute to lists, requires prior specic permission and/or a
fee. Request permissions from permissions@acm.org.
ICBIP 2022, August 19–21, 2022, Suzhou, China
© 2022 Association for Computing Machinery.
ACM ISBN 978-1-4503-9669-1/22/08.. . $15.00
https://doi.org/10.1145/3563737.3563759
ACM Reference Format:
SHUANGMING MAO and SEIICHIRO KAMATA. 2022. MRI Super-Resolution
using Implicit Neural Representation with Frequency Domain Enhancement.
In 2022 7th International Conference on Biomedical Signal and Image Pro-
cessing (ICBIP) (ICBIP 2022), August 19–21, 2022, Suzhou, China. ACM, New
York, NY, USA, 7 pages. https://doi.org/10.1145/3563737.3563759
EDSR
mDCSRN
SRCNN VDSR
ours
1.0
0.1
Figure 1: Visual comparison with error maps of low-
frequency domain for dierent methods on dataset IXI-T2
1 INTRODUCTION
High Resolution (HR) Magnetic Resonance Imaging (MRI) provides
accurate structural information and textures, which is useful to clin-
ical diagnosis and quantitative image analyses. However, obtaining
HR MR images remains a dicult challenge since it takes a lot of
scan time, lower peak signal-to-noise ratio (PSNR) and smaller spa-
tial converge [
28
]. Single Image Super-Resolution (SISR) becomes
a promising technique to address this issue. SISR can produce HR
image from a single Low-Resolution (LR) image, which can signi-
cantly reduce the scanning time. Therefore, improve the ability of
SISR is urgently needed for decades.
Deep convolutional neural networks (CNNs) show its outstand-
ing performance for Image Super-Resolution [
4
,
14
,
17
,
18
,
40
,
45
].
There are 3 major directions to improve the performance for CNN-
based SR: 1. Design a more complicated and deeper architectures. 2.
Provide additional information with help of extra inputs. 3. Convert
SR tasks to another tasks. For instance, VDSR [
14
] increase the
network depth with Residual Learning [
8
]. SRNTT [
45
] provides an
extra HR image as input and design a Neural Texture Transformer
module to integrate texture information to CNN. [
21
] utilizes a
pre-trained StyleGAN [
13
] to generate HR images, which converts
SR problem to generative problem and produces promising results
in Face SR.
However, previous CNN-based works ignore the characteris-
tics of MR images and treat MR images as same as natural images.
First, MRI Super-Resolution should keep the original information
unchanged. Fig 1 shows the visual comparison for error maps of
k-space between our method and another methods. As shown in
30