Study finds no overall winner between 3D and 2.5D U-Nets for diffusion-based MRI super-resolution
An arXiv abstract reports that two U-Net backbones for an elucidated diffusion MRI super-resolution model perform comparably, with each leading under different test conditions.
In plain words
Super-resolution means using software to sharpen low-resolution MRI scans instead of buying stronger scanners. The authors trained a diffusion model with two different network designs and found neither is universally better: results depend on the test data and degradation type.
What was announced
The study, published as an arXiv abstract, compares a 3D convolutional U-Net that processes volumetric patches against a 2.5D slice-conditioned U-Net that super-resolves each slice while using its two neighbouring slices as context. Both backbones were trained from scratch on 500 subjects from the NKI cohort of FOMO60K, with checkpoint selection on a 6-subject validation set and evaluation on 70 held-out subjects against EDSR, Swin2SR and Res-SRDiff baselines trained on the same data.
On the degradation used in training, the two architectures tied at 37.72 dB and 37.70 dB PSNR, 0.2 dB ahead of EDSR. Under k-space truncation, EDSR matched the diffusion models, and with added noise the 2.5D model was most accurate. On 70 external IXI subjects, the 3D model led, 0.4 dB above the best baseline and 0.9 dB above the 2.5D model in foreground PSNR. The authors conclude neither architecture is better overall.
What it does not show
- The evidence is an abstract only; full methods, additional results and peer review status are not available in the supplied passage.
- Results cover brain MRI at 2x in-plane super-resolution on NKI and IXI cohorts and may not extend to other anatomies, resolutions or degradation types.
- The reported comparisons come from the authors' own evaluation and have not been independently confirmed.
Sources
Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 17 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.