Research map / Label-efficient and robust learning / Recognition and long-tail
Robustness of Deep Learning for Accelerated MRI: Benefits of Diverse Training Data
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.
Shows that training one accelerated-MRI reconstruction model on diverse data (multiple anatomies, contrasts, field strengths, ~413k slices from 13 datasets) matches separate per-distribution models in-distribution and improves out-of-distribution robustness, while noting distributional overfitting under long training.
Problem
Deep learning MRI reconstruction models degrade under distribution shifts (anatomy, contrast, scanner, field strength), and it is unclear whether training on diverse data compromises in-distribution performance or helps robustness.
Approach
Training a single model on a union of diverse MRI distributions matches separately trained models in-distribution and yields effective robustness (above the in-distribution vs. out-of-distribution trend line), with robustness correlated with CLIP-feature nearest-neighbor similarity between training and test sets; early stopping mitigates distributional overfitting.
Key results
- Models trained jointly on two distributions (e.g., knees and brains) match separately trained models in-distribution, and this holds for U-net, ViT and VarNet.
- A VarNet trained on the 13-dataset collection almost closes the out-of-distribution gap on CC-359 sagittal relative to a model trained on that dataset itself.
- Training on diverse data gives effective robustness: models trained on diverse sets lie above the in-distribution vs. out-of-distribution trend line, with robustness correlating with CLIP nearest-neighbor similarity.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train on a diverse collection of datasets instead of fastMRI alone | U-net · Out-of-distribution MRI datasets (CC-359 sagittal, NYU, Stanford 2D, M4Raw GRE) | LPIPS/DISTS: CC-359 sagittal: 0.294/0.185 -> 0.247/0.149; NYU: 0.226/0.128 -> 0.221/0.128; Stanford 2D: 0.260/0.163 -> 0.257/0.154; M | Helped |
| Train on a diverse collection of datasets instead of fastMRI alone | ViT · Out-of-distribution MRI datasets (CC-359 sagittal, NYU, Stanford 2D, M4Raw GRE) | LPIPS/DISTS: CC-359 sagittal: 0.264/0.168 -> 0.228/0.149; NYU: 0.254/0.150 -> 0.219/0.128; Stanford 2D: 0.258/0.168 -> 0.243/0.154; M | Helped |
| Train jointly on 2-fold, 4-fold, 8-fold, and 16-fold acceleration | U-net · fastMRI PDFS knee, 3-fold acceleration | SSIM: 0.906 -> 0.912 | Helped |
| Train jointly on 2-fold, 4-fold, 8-fold, and 16-fold acceleration | U-net · fastMRI PDFS knee, 2-fold, 4-fold, 8-fold, and 16-fold acceleration | SSIM: 2-fold: 0.945 -> 0.944; 4-fold: 0.903 -> 0.902; 8-fold: 0.867 -> 0.866; 16-fold: 0.828 -> 0.829 | No clear effect |
| Use one regularization weight across distributions rather than tuning it separately | Wavelet l1 least-squares · fastMRI knee, PD Knee Skyra 3.0T and PDFS Knee Aera 1.5T | SSIM: P: 0.788 -> 0.792; Q: 0.602 -> 0.609 | Mixed |
| Train on diverse healthy-subject data without pathology examples | U-net, ViT, and VarNet · fastMRI brain pathology regions | SSIM | No clear effect |
| Stop training early rather than continue to 60 epochs | U-net, ViT, and VarNet · Out-of-distribution MRI, including fastMRI T2 brain to fastMRI knee | SSIM | Helped |
| Use SGD instead of Adam (ablation) | U-net · fastMRI brain and fastMRI knee | SSIM | No clear effect |
| Train jointly on P and Q rather than train separate models on each distribution | U-net, ViT, and VarNet · fastMRI anatomy, contrast, and magnetic-field distributions | SSIM | No clear effect |
| Train one model on skewed P+Q data, with P about 10 times smaller than Q | U-net · fastMRI anatomy, contrast, and magnetic-field distributions | SSIM | No clear effect |
Surprising findings
- Models trained on images without pathologies reconstruct pathologies as accurately as models trained with pathologies.
- Out-of-distribution performance can drop with prolonged training while in-distribution performance keeps improving slightly (distributional overfitting).
Limitations
- Training on a diverse dataset often does not close the gap to the idealized model trained on the out-of-distribution data.
- Collecting large and diverse datasets may be difficult or expensive in practice.
- Results focus on 4-fold Cartesian 2D multi-coil MRI and specific public datasets.
Experimental setup
Datasets: fastMRI knee fastMRI brain fastMRI prostate M4Raw SKM-TEA Stanford 3D Stanford 2D 7T database CC-359 NYU data fastMRI+
Benchmarks: fastMRI knee validation fastMRI brain test CC-359 sagittal Stanford 2D M4Raw GRE NYU data
Metrics: SSIM LPIPS DISTS effective robustness
Baselines: U-net ViT end-to-end VarNet l1-regularized least squares
Builds on: fastMRI CLIP effective robustness
Closest papers by meaning
- Improving Deep Learning for Accelerated MRI With Data Filtering (NeurIPS 2025 · similarity 0.90)
- Measuring Robustness in Deep Learning Based Compressive Sensing (ICML 2021 · similarity 0.89)
- Data augmentation for deep learning based accelerated MRI reconstruction with limited data (ICML 2021 · similarity 0.89)
- Efficient Noise Calculation in Deep Learning-based MRI Reconstructions (ICML 2025 · similarity 0.87)
- Dynamic Sparse Training versus Dense Training: The Unexpected Winner in Image Corruption Robustness (ICLR 2025 · similarity 0.85)
- Fast MRI for All: Bridging Access Gaps by Training without Raw Data (NeurIPS 2025 · similarity 0.84)
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.