Research map / Image generation / Super-resolution
Rethinking Diffusion Model for Multi-Contrast MRI Super-Resolution
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DiffMSR, a diffusion model run in a compact low-dimensional latent space (4 steps) that generates prior features to guide a large-window Transformer decoder (PLWformer) for multi-contrast MRI super-resolution.
Problem
DM-based MRI super-resolution needs many iterations, which is computationally expensive, and often produces artifacts and misaligned, distorted anatomical details.
Approach
Apply the diffusion process in a highly compact latent space learned by a Prior Extraction module from the HR image, so only a few denoising steps are needed; a Prior-Guide Large Window Transformer (PLWformer) with spatially reduced keys/values decodes the prior together with T1 reference features to produce undistorted SR images.
Key results
- On FastMRI Knee at 4x, DiffMSR reaches 33.78 dB PSNR / 0.8765 SSIM, versus 32.24 dB for DisC-Diff (the best baseline shown in Table 1).
- DiffMSR uses 6.8M parameters, 58.6 GFLOPs and 39 ms inference, against 86.1M parameters, 461 GFLOPs and 1455 ms for DisC-Diff.
- Ablation: removing the reference image drops PSNR to 32.56 dB and removing the prior drops it to 31.51 dB, compared with 33.78 dB for the full model.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace DisC-Diff with DiffMSR | MRI super-resolution · FastMRI | PSNR (dB): 32.24 -> 33.78 | Helped |
| replace DisC-Diff with DiffMSR | MRI super-resolution · Clinical Brain | PSNR (dB): 32.85 -> 34.12 | Helped |
| replace DisC-Diff with DiffMSR | MRI super-resolution · Clinical Tumor | PSNR (dB): 32.64 -> 33.85 | Helped |
| replace DisC-Diff with DiffMSR | MRI super-resolution · Clinical Pelvic | PSNR (dB): 34.26 -> 35.47 | Helped |
| replace DisC-Diff with DiffMSR | MRI super-resolution · FastMRI | SSIM: 0.8530 -> 0.8765 | Helped |
| replace DisC-Diff with DiffMSR | MRI super-resolution · Clinical Brain | SSIM: 0.9285 -> 0.9362 | Helped |
| replace DisC-Diff with DiffMSR | MRI super-resolution · Clinical Tumor | SSIM: 0.9187 -> 0.9274 | Helped |
| replace DisC-Diff with DiffMSR | MRI super-resolution · Clinical Pelvic | SSIM: 0.9388 -> 0.9607 | Helped |
| add the reference image (ablation) | DiffMSR · FastMRI | PSNR: 32.56 -> 33.78 | Helped |
| add prior extraction and guidance (ablation) | DiffMSR · FastMRI | PSNR: 31.51 -> 33.78 | Helped |
| jointly train the diffusion model and PLWformer in stage two (ablation) | DiffMSR · FastMRI | PSNR: 32.35 -> 33.78 | Helped |
| add data consistency loss (ablation) | DiffMSR · FastMRI | PSNR: 32.92 -> 33.78 | Helped |
Surprising findings
- With only one diffusion step (T=1) the DM fails to generate reasonable prior knowledge, while the PSNR curve flattens by T=4.
Limitations
- Evaluated only at 4x upscaling.
- Clinical datasets are private, and the public data is a limited subset (1600 slices).
Experimental setup
Datasets: FastMRI Knee clinical brain MRI (T1 reference, T2-FLAIR target) clinical pelvic MRI (T1 reference, T2 target) clinical tumor MRI
Benchmarks: FastMRI Knee clinical brain clinical tumor clinical pelvic
Metrics: PSNR SSIM parameters FLOPs inference speed
Baselines: MCSR MINet MASA WavTrans McMRSR MC-VarNet DisC-Diff
Builds on: DDPM Swin Transformer DisC-Diff
Closest papers by meaning
- FaithDiff: Unleashing Diffusion Priors for Faithful Image Super-resolution (CVPR 2025 · similarity 0.87)
- SinSR: Diffusion-Based Image Super-Resolution in a Single Step (CVPR 2024 · similarity 0.87)
- Self-Supervised Diffusion MRI Denoising via Iterative and Stable Refinement (ICLR 2025 · similarity 0.87)
- Effective Diffusion Transformer Architecture for Image Super-Resolution (AAAI 2025 · similarity 0.87)
- Unleashing the Power of One-Step Diffusion based Image Super-Resolution via a Large-Scale Diffusion Discriminator (NeurIPS 2025 · similarity 0.87)
- Autoregressive Image Diffusion: Generation of Image Sequence and Application in MRI (NeurIPS 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.