Skip to content
DraconiticResearch
Search Sources API
197,787 studies31 sourcesUpdated 5 min ago
Peer-reviewedCC BY 4.0Imaging Neuroscience

SRNR: Deep learning-based MRI super-resolution using noisy high-resolution reference data

  • Muye Zhang1
  • Hongjia Yang1
  • Zihan Li1
  • Jiaxin Xiao1
  • Wanqing Liu2
  • Yuhang He1
  • Ting Yin3
  • Jialan Zheng1
  • Mingxuan Liu1
  • Wei Liu4
  • Congyu Liao5
  • Berkin Bilgic6,7,8
6 more authors
  • Jonathan R. Polimeni6,7,8
  • Susie Y. Huang6,7,8
  • Yi Liao2
  • Ziyu Li9
  • Haibo Qu2
  • Qiyuan Tian1
9 affiliations
  1. 1School of Biomedical Engineering, Tsinghua University, Beijing, China
  2. 2Department of Radiology, West China Second University Hospital, Sichuan University, Chengdu, China
  3. 3MR Research Collaborations, Siemens Healthineers Ltd., Chengdu, China
  4. 4Siemens Healthineers AG, Erlangen, Germany
  5. 5Department of Radiology and Biomedical Imaging, University of California San Francisco, San Francisco, CA, United States
  6. 6Athinoula A. Martinos Center for Biomedical Imaging, Department of Radiology, Massachusetts General Hospital, Charlestown, MA, United States
  7. 7Department of Radiology, Harvard Medical School, Boston, MA, United States
  8. 8Harvard-MIT Program in Health Sciences and Technology, Massachusetts Institute of Technology, Cambridge, MA, United States
  9. 9Oxford Centre for Integrative Neuroimaging, FMRIB, Nuffield Department of Clinical Neurosciences, University of Oxford, Oxford, United Kingdom
Read full textDOICite

Abstract

Deep learning-based super-resolution (SR) techniques offer a powerful tool to enhance the spatial resolution of magnetic resonance imaging (MRI) without increasing scan time. However, most supervised SR methods rely on high-resolution, high-signal-to-noise ratio (SNR) reference data for training, which are difficult to acquire in many clinical settings, such as in time-constrained pediatric imaging. Inspired by the Noise2Noise principle, we propose SRNR, a new training paradigm that enables supervised SR using noisy high-resolution images as references. SRNR was systematically evaluated on both pediatric and adult simulated data across different noise levels of the reference data. SRNR generated super-resolved image volumes with high similarity to the high-resolution references (PSNR: 32.5–35.2 dB) and improved downstream image analysis performance, including tissue segmentation (white matter Dice coefficient: 0.951–0.960) and cortical surface reconstruction (cortical thickness absolute difference: 0.200–0.255 mm). Furthermore, SRNR was shown to be effective on clinically acquired pediatric brain MRI data, achieving improvements consistent with those observed in the simulated experiments. These results demonstrate that SRNR substantially reduces reliance on clean, high-SNR training data and enables effective SR under clinically realistic acquisition constraints, thereby improving the feasibility of deploying deep learning-based SR in pediatric MRI and related applications.

Related

  1. Super-resolution imaging flow cytometry
    Y. Luo, T. Ding, Y. Liu +7
    bioRxivPreprintAll rights reservedFull textSystems & synthetic biologybioRxiv
  2. Multiscale, Survey-and-Target Optical Super-Resolution Histology
    Kidan Tadesse, Oliver Lee, Hajun Yoo +10
    bioRxivPreprintAll rights reservedFull textSystems & synthetic biologybioRxivPMC
  3. Ratio-preserving super-resolution FLIM for quantitative live-cell lifetime imaging
    C. Li, L. Li, Y. Cheng +13
    bioRxivPreprintAll rights reservedFull textSystems & synthetic biologybioRxiv
  4. ASTRA enables zero-shot super-resolution of spatial transcriptomics across capture geometries
    J. Ye, Q. Cui
    bioRxivPreprintCC BY-NC-NDFull textBioinformatics & computational biologybioRxiv
  5. RhoBAST-FLIM enables high-fidelity and lifetime-assisted super-resolution RNA imaging
    X.-X. Chen, B. Bozkurt, M. Sunbul
    bioRxivPreprintCC BYFull textSystems & synthetic biologybioRxiv

Cite

BibTeXRIS
BibTeX
@article{zhang2026srnr,
  title = {{SRNR: Deep learning-based MRI super-resolution using noisy high-resolution reference data}},
  author = {Muye Zhang and Hongjia Yang and Zihan Li and Jiaxin Xiao and Wanqing Liu and Yuhang He and Ting Yin and Jialan Zheng and Mingxuan Liu and Wei Liu and Congyu Liao and Berkin Bilgic and Jonathan R. Polimeni and Susie Y. Huang and Yi Liao and Ziyu Li and Haibo Qu and Qiyuan Tian},
  journal = {Imaging Neuroscience},
  year = {2026},
  month = nov,
  doi = {10.1162/imag.a.1433},
  url = {https://doi.org/10.1162/imag.a.1433},
  note = {CC BY 4.0},
}
RIS
TY  - JOUR
TI  - SRNR: Deep learning-based MRI super-resolution using noisy high-resolution reference data
AU  - Muye Zhang
AU  - Hongjia Yang
AU  - Zihan Li
AU  - Jiaxin Xiao
AU  - Wanqing Liu
AU  - Yuhang He
AU  - Ting Yin
AU  - Jialan Zheng
AU  - Mingxuan Liu
AU  - Wei Liu
AU  - Congyu Liao
AU  - Berkin Bilgic
AU  - Jonathan R. Polimeni
AU  - Susie Y. Huang
AU  - Yi Liao
AU  - Ziyu Li
AU  - Haibo Qu
AU  - Qiyuan Tian
PY  - 2026
DA  - 2026/11/03
JO  - Imaging Neuroscience
DO  - 10.1162/imag.a.1433
UR  - https://doi.org/10.1162/imag.a.1433
AB  - Deep learning-based super-resolution (SR) techniques offer a powerful tool to enhance the spatial resolution of magnetic resonance imaging (MRI) without increasing scan time. However, most supervised SR methods rely on high-resolution, high-signal-to-noise ratio (SNR) reference data for training, which are difficult to acquire in many clinical settings, such as in time-constrained pediatric imaging. Inspired by the Noise2Noise principle, we propose SRNR, a new training paradigm that enables supervised SR using noisy high-resolution images as references. SRNR was systematically evaluated on both pediatric and adult simulated data across different noise levels of the reference data. SRNR generated super-resolved image volumes with high similarity to the high-resolution references (PSNR: 32.5–35.2 dB) and improved downstream image analysis performance, including tissue segmentation (white matter Dice coefficient: 0.951–0.960) and cortical surface reconstruction (cortical thickness absolute difference: 0.200–0.255 mm). Furthermore, SRNR was shown to be effective on clinically acquired pediatric brain MRI data, achieving improvements consistent with those observed in the simulated experiments. These results demonstrate that SRNR substantially reduces reliance on clean, high-SNR training data and enables effective SR under clinically realistic acquisition constraints, thereby improving the feasibility of deploying deep learning-based SR in pediatric MRI and related applications.
SN  - 2837-6056
ER  -