Research map / Label-efficient and robust learning / Recognition and long-tail
Improving Deep Learning for Accelerated MRI With Data Filtering
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.
Assembles 1.1M-slice raw k-space corpus from 18 public sources and shows that DreamSim-based alignment filtering of training data (weighted variant best) modestly improves VarNet and diffusion MRI reconstruction across 48 curated test sets.
Problem
Deep learning for accelerated MRI reconstruction has focused on architectures, while training data curation for reconstruction networks is largely unstudied; it is unclear how to construct training sets for high-performing, robust models.
Approach
Adapts retrieval-based data filtering from vision-language work to MRI: embed MVUE target images with DreamSim, retrieve k-nearest-neighbor training slices to a curated validation set, and optionally weight sampling by retrieval frequency (weighted alignment filtering), alongside simple energy and edge-density heuristics.
Key results
- Weighted alignment filtering achieves the best mean SSIM over 48 test sets among filters, improving on 46 of 48 test sets at 4-fold acceleration with 120k unfiltered slices.
- At 4-fold acceleration with 120k unfiltered slices and 1% in-distribution data, weighted alignment filtering reaches 40.60 dB mean PSNR vs 39.99 dB with no filtering (40k vs 120k slices).
- Filtering gains are comparable to a 3-fold increase in unfiltered training data, and filtering helps most when in-distribution data fraction is low.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply weighted DreamSim alignment filtering instead of no filtering | VarNet · Mean over 48 datasets, 4-fold acceleration | PSNR [dB]: 39.99 -> 40.60 | Helped |
| Apply weighted DreamSim alignment filtering instead of no filtering | VarNet · Mean over 48 datasets, 4-fold acceleration | LPIPS: 0.111 -> 0.108 | Helped |
| Use weighted alignment filtering instead of unweighted alignment filtering (ablation) | VarNet · Mean over 48 datasets, 4-fold acceleration | PSNR [dB]: 40.38 -> 40.60 | Helped |
| Add heuristic filtering before alignment filtering (ablation) | VarNet · Mean over 48 datasets, 4-fold acceleration | PSNR [dB]: 40.38 -> 40.42 | No clear effect |
| Use DreamSim retrieval instead of Euclidean pixel-distance retrieval (ablation) | VarNet · All test sets, 8-fold acceleration | SSIM gain over no filtering: 0.0 -> +0.006 | Helped |
| Use Euclidean pixel-distance retrieval instead of DreamSim retrieval (ablation) | VarNet · Cardiac, 8-fold acceleration | SSIM gain over no filtering: --0.010 | Hurt |
| Apply weighted alignment filtering instead of no filtering | U-net · Mean over 48 datasets, 4-fold acceleration | SSIM: 0.910 -> 0.913 | Helped |
| Apply weighted alignment filtering instead of no filtering | ViT · Mean over 48 datasets, 4-fold acceleration | SSIM: 0.920 -> 0.923 | Helped |
| Apply weighted alignment filtering instead of no filtering | VarNet · 48 test sets, 4-fold acceleration | SSIM: 46 out of 48 | Helped |
| Apply weighted alignment filtering when the unfiltered set has little in-distribution data (ablation) | VarNet · In-distribution test data, 8-fold acceleration | performance | Helped |
| Apply weighted alignment filtering when the unfiltered set is completely in-distribution (ablation) | VarNet · In-distribution test data, 8-fold acceleration | performance | No clear effect |
| Apply weighted alignment filtering when no auxiliary data is used (ablation) | VarNet · In-distribution test data, 8-fold acceleration | performance | Hurt |
Surprising findings
- Filtering can hurt performance when the unfiltered training set consists entirely of in-distribution data.
- A pixel-based (Euclidean) retrieval metric yields no average improvement across all test sets, while DreamSim retrieval does.
Limitations
- Gains are modest on average (e.g., ~0.01 SSIM), and the authors note they are smaller than in language and vision-language domains.
- Focused only on 2D accelerated MRI; 3D, motion-compensated, and non-MRI reconstruction are not studied.
- Curated datasets cannot be hosted due to licensing; users must reconstruct from public sources. Filtering relies on a curated validation set resembling the test distribution.
Experimental setup
Datasets: fastMRI knee fastMRI brain CMRxRecon2023 M4Raw SKM-TEA AHEAD fastMRI breast Lung 3D UTE Chirp 3D Extreme MRI Fruits Phantom Heart T2-mapping
Benchmarks: 48 curated accelerated 2D MRI test sets (21 in-distribution, 27 out-of-distribution) 3D MRI test sets (AHEAD, SKM-TEA, Stanford-3D, CC359)
Metrics: SSIM PSNR LPIPS Frechet DreamSim Distance
Baselines: No filtering Heuristic filtering Alignment filtering Pixel-based retrieval (Euclidean distance) U-net Vision Transformer
Builds on: VarNet DreamSim DDPM decomposed diffusion sampling
Closest papers by meaning
- Robustness of Deep Learning for Accelerated MRI: Benefits of Diverse Training Data (ICML 2024 · similarity 0.90)
- Efficient Noise Calculation in Deep Learning-based MRI Reconstructions (ICML 2025 · similarity 0.88)
- Progressive Divide-and-Conquer via Subsampling Decomposition for Accelerated MRI (CVPR 2024 · similarity 0.87)
- Fast MRI for All: Bridging Access Gaps by Training without Raw Data (NeurIPS 2025 · similarity 0.86)
- UGoDIT: Unsupervised Group Deep Image Prior Via Transferable Weights (NeurIPS 2025 · similarity 0.86)
- Breaking Information Isolation: Accelerating MRI via Inter-sequence Mapping and Progressive Masking (AAAI 2025 · similarity 0.85)
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.