Research map / Image generation / Image restoration
Fast MRI for All: Bridging Access Gaps by Training without Raw 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.
Proposes CUPID, which trains physics-driven deep learning MRI reconstruction using only clinical parallel-imaging DICOM images by combining a reweighted-l1 compressibility loss with a perturbation-based parallel imaging fidelity loss.
Problem
PD-DL MRI reconstruction training normally requires raw k-space data, which is unavailable at many rural or non-specialized scanners that only export reconstructed DICOM images, limiting generalization fine-tuning.
Approach
Train an unrolled PD-DL network on sub-sampled, parallel-imaging-reconstructed DICOM images using a compressibility (reweighted wavelet l1) loss plus an equivariance-based parallel imaging fidelity term driven by added known perturbations (letters, shapes) that are resolvable by parallel imaging, avoiding the all-zeros collapse without raw k-space.
Key results
- On retrospective R=4 database training, CUPID reaches 38.82 dB / 0.952 on coronal PD knee, close to supervised PD-DL (40.44 dB / 0.964) and above CS (36.92 dB), EI, DDS, and ScoreMRI.
- CUPID outperforms DDS on axial FLAIR brain (36.49 dB / 0.957 vs 36.18 / 0.950) and axial T2 (35.31 dB / 0.921 vs 35.16 / 0.914).
- Zero-shot CUPID runtime is about 7-8 minutes per slice with K=6 perturbations, and radiologists judged its outputs diagnostically similar to references on FastMRI+ pathology cases.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train the PD-DL model with CUPID using sub-sampled DICOM images instead of raw k-space | CUPID · Cor PD, Knee MRI R=4 | PSNR: 35.51 -> 38.82 | Helped |
| Train the PD-DL model with CUPID using sub-sampled DICOM images instead of raw k-space | CUPID · Cor PD-FS, Knee MRI R=4 | PSNR: 29.48 -> 35.04 | Helped |
| Train the PD-DL model with CUPID using sub-sampled DICOM images instead of raw k-space | CUPID · Ax FLAIR, Brain MRI R=4 | PSNR: 31.97 -> 36.49 | Helped |
| Train the PD-DL model with CUPID using sub-sampled DICOM images instead of raw k-space | CUPID · Ax T2, Brain MRI R=4 | PSNR: 31.02 -> 35.31 | Helped |
| Use CUPID rather than supervised PD-DL | PD-DL reconstruction · Cor PD, Knee MRI R=4 | PSNR: 40.44 -> 38.82 | Hurt |
| Use CUPID rather than SSDU | PD-DL reconstruction · Cor PD-FS, Knee MRI R=4 | SSIM: 0.892 -> 0.880 | No clear effect |
| Use six perturbations rather than fewer perturbations (ablation) | CUPID zero-shot training · Cor PD knee MRI R=4 | reconstruction quality | Helped |
| Increase perturbation count beyond six (ablation) | CUPID zero-shot training · Cor PD knee MRI R=4 | reconstruction quality | No clear effect |
| Remove parallel imaging fidelity by setting lambda=0 (ablation) | CUPID · MRI reconstruction | image fidelity | Hurt |
| Remove compressibility by using only parallel imaging fidelity (ablation) | CUPID · MRI reconstruction | noise amplification | Hurt |
| Use lambda in the mid-range rather than either loss term alone (ablation) | CUPID · MRI reconstruction | reconstruction quality | Helped |
| Use random uniform rather than equidistant undersampling at R=4 (ablation) | CUPID zero-shot reconstruction · FastMRI knee and brain | reconstruction quality | No clear effect |
Surprising findings
- Using only the parallel imaging fidelity term (lambda to infinity) amplifies noise, and using only the compressibility term yields overly smooth images, so both terms are needed.
- Higher PSNR/SSIM for ScoreMRI-style blurring is noted as favoring smoother outputs, illustrating the perception-distortion trade-off.
Limitations
- Performance degrades at R=8 with visible residual artifacts, and it trails supervised methods on some datasets like PD-FS.
- Coil sensitivities must be estimated when calibration scans are unavailable, which is outside the study's scope.
- Vendor filtering or processing (e.g., implicit intensity correction, deidentification) can violate the assumed parallel imaging forward model.
Experimental setup
Datasets: fastMRI knee (coronal PD, PD-FS) fastMRI brain (axial T2, FLAIR) FastMRI+ prospective 7T MP2RAGE
Benchmarks: fastMRI knee fastMRI brain FastMRI+
Metrics: PSNR SSIM
Baselines: Supervised PD-DL SSDU Equivariant imaging ScoreMRI DDS CG-SENSE Compressed sensing ZS-SSDU
Builds on: equivariant imaging self-supervision via data undersampling (SSDU) variable splitting with quadratic penalty (VS-QP) compressibility-inspired unsupervised learning
Closest papers by meaning
- Improving Deep Learning for Accelerated MRI With Data Filtering (NeurIPS 2025 · similarity 0.86)
- A Unified Model for Compressed Sensing MRI Across Undersampling Patterns (CVPR 2025 · similarity 0.86)
- Efficient Noise Calculation in Deep Learning-based MRI Reconstructions (ICML 2025 · similarity 0.85)
- D2SA: Dual-Stage Distribution and Slice Adaptation for Efficient Test-Time Adaptation in MRI Reconstruction (NeurIPS 2025 · similarity 0.84)
- UGoDIT: Unsupervised Group Deep Image Prior Via Transferable Weights (NeurIPS 2025 · similarity 0.84)
- Progressive Divide-and-Conquer via Subsampling Decomposition for Accelerated MRI (CVPR 2024 · 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.