SRMRI: A Diffusion-Based Super-resolution Framework and Open Dataset for Blind MRI Super-Resolution

Arpan Poudel, Mamata Shrestha, Nian Wang, Ukash Nakarmi
Proceedings of The 8th International Conference on Medical Imaging with Deep Learning, PMLR 301:1326-1341, 2026.

Abstract

Existing deep learning methods for medical image super-resolution (SR) often rely on paired datasets generated by simulating low-resolution (LR) images from corresponding high-resolution (HR) scans, which can introduce biases and degrade real-world performance. To overcome these limitations, we present an unsupervised approach based on a score-based diffusion model that does not require paired training data. We train a score-based diffusion model using denoising score matching on HR Magnetic Resonance Imaging (MRI) scans, then perform iterative refinement with a stochastic differential equation (SDE) solver while enforcing data consistency from LR scans. Our method provides faster sampling compared to existing generative approaches and achieves competitive results on key metrics, though it does not surpass fully supervised baselines in PSNR and SSIM. Notably, while supervised models often report higher numerical metrics, we observe that they can produce suboptimal reconstructions due to their reliance on fixed upscaling kernels. Finally, we introduce the SRMRI dataset, containing LR and HR images obtained from scanner for training and evaluating MR image super-resolution models. Code and dataset are available at: https://github.com/arpanpoudel/SRMRI

Cite this Paper

BibTeX
@InProceedings{pmlr-v301-poudel26a, title = {SRMRI: A Diffusion-Based Super-resolution Framework and Open Dataset for Blind MRI Super-Resolution}, author = {Poudel, Arpan and Shrestha, Mamata and Wang, Nian and Nakarmi, Ukash}, booktitle = {Proceedings of The 8th International Conference on Medical Imaging with Deep Learning}, pages = {1326--1341}, year = {2026}, editor = {Tasdizen, Tolga and Elhabian, Shireen and Summers, Ronald and Chen, Chen and Koch, Lisa and Zhuang, Yan}, volume = {301}, series = {Proceedings of Machine Learning Research}, month = {09--11 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v301/main/assets/poudel26a/poudel26a.pdf}, url = {https://proceedings.mlr.press/v301/poudel26a.html}, abstract = {Existing deep learning methods for medical image super-resolution (SR) often rely on paired datasets generated by simulating low-resolution (LR) images from corresponding high-resolution (HR) scans, which can introduce biases and degrade real-world performance. To overcome these limitations, we present an unsupervised approach based on a score-based diffusion model that does not require paired training data. We train a score-based diffusion model using denoising score matching on HR Magnetic Resonance Imaging (MRI) scans, then perform iterative refinement with a stochastic differential equation (SDE) solver while enforcing data consistency from LR scans. Our method provides faster sampling compared to existing generative approaches and achieves competitive results on key metrics, though it does not surpass fully supervised baselines in PSNR and SSIM. Notably, while supervised models often report higher numerical metrics, we observe that they can produce suboptimal reconstructions due to their reliance on fixed upscaling kernels. Finally, we introduce the SRMRI dataset, containing LR and HR images obtained from scanner for training and evaluating MR image super-resolution models. Code and dataset are available at: https://github.com/arpanpoudel/SRMRI} }
Endnote
%0 Conference Paper %T SRMRI: A Diffusion-Based Super-resolution Framework and Open Dataset for Blind MRI Super-Resolution %A Arpan Poudel %A Mamata Shrestha %A Nian Wang %A Ukash Nakarmi %B Proceedings of The 8th International Conference on Medical Imaging with Deep Learning %C Proceedings of Machine Learning Research %D 2026 %E Tolga Tasdizen %E Shireen Elhabian %E Ronald Summers %E Chen Chen %E Lisa Koch %E Yan Zhuang %F pmlr-v301-poudel26a %I PMLR %P 1326--1341 %U https://proceedings.mlr.press/v301/poudel26a.html %V 301 %X Existing deep learning methods for medical image super-resolution (SR) often rely on paired datasets generated by simulating low-resolution (LR) images from corresponding high-resolution (HR) scans, which can introduce biases and degrade real-world performance. To overcome these limitations, we present an unsupervised approach based on a score-based diffusion model that does not require paired training data. We train a score-based diffusion model using denoising score matching on HR Magnetic Resonance Imaging (MRI) scans, then perform iterative refinement with a stochastic differential equation (SDE) solver while enforcing data consistency from LR scans. Our method provides faster sampling compared to existing generative approaches and achieves competitive results on key metrics, though it does not surpass fully supervised baselines in PSNR and SSIM. Notably, while supervised models often report higher numerical metrics, we observe that they can produce suboptimal reconstructions due to their reliance on fixed upscaling kernels. Finally, we introduce the SRMRI dataset, containing LR and HR images obtained from scanner for training and evaluating MR image super-resolution models. Code and dataset are available at: https://github.com/arpanpoudel/SRMRI
APA
Poudel, A., Shrestha, M., Wang, N. & Nakarmi, U.. (2026). SRMRI: A Diffusion-Based Super-resolution Framework and Open Dataset for Blind MRI Super-Resolution. Proceedings of The 8th International Conference on Medical Imaging with Deep Learning, in Proceedings of Machine Learning Research 301:1326-1341 Available from https://proceedings.mlr.press/v301/poudel26a.html.

Related Material