Research map / Image generation / Image restoration
A Unified Model for Compressed Sensing MRI Across Undersampling Patterns
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 a discretization-agnostic neural operator (UDNO built on DISCO convolutions) unrolled network for compressed sensing MRI that works across undersampling patterns and resolutions.
Problem
Deep learning MRI reconstruction models are tied to a fixed discretization, so separate models are needed for each undersampling pattern and output resolution, which is impractical clinically.
Approach
Replace CNN priors in an unrolled network with neural operators built from DISCO local-integral convolutions applied in both k-space (NOk) and image space (NOi), giving resolution-agnostic kernels that also enable zero-shot super-resolution and extended field of view.
Key results
- Achieves 39.14 dB PSNR / 0.9219 SSIM on fastMRI knee at 4x equispaced vs. 38.33 dB / 0.9048 for E2E-VN.
- Averages 4.17 dB PSNR and 8.4% SSIM gain over E2E-VN across undersampling patterns on fastMRI knee; 3.2 dB higher PSNR across rates on knee.
- Inference time 0.158 s vs. 84-96 s for diffusion baselines (about 600x faster) on an A100.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace CNN-based E2E-VN priors with resolution-agnostic neural operators | MRI reconstruction on fastMRI knee · fastMRI knee, 4× equispaced | PSNR (dB): 38.33±3.06 -> 39.14±2.93 | Helped |
| Replace CNN-based E2E-VN priors with resolution-agnostic neural operators | MRI reconstruction on fastMRI knee · fastMRI knee, 4× equispaced | SSIM: 0.9048±0.0732 -> 0.9219±0.0724 | Helped |
| Replace CNN-based E2E-VN priors with resolution-agnostic neural operators | MRI reconstruction on fastMRI brain · fastMRI brain, 4× equispaced | PSNR (dB): 38.06±2.70 -> 38.82±2.77 | Helped |
| Replace CNN-based E2E-VN priors with resolution-agnostic neural operators | MRI reconstruction on fastMRI brain · fastMRI brain, 4× equispaced | SSIM: 0.9620±0.0107 -> 0.9621±0.0086 | No clear effect |
| Use neural operators instead of E2E-VN++ | MRI reconstruction on fastMRI knee · fastMRI knee, 4× equispaced | PSNR (dB): 37.50 ± 2.79 -> 37.40 ± 2.61 | No clear effect |
| Use neural operators instead of E2E-VN++ | MRI reconstruction on fastMRI knee · fastMRI knee, 4× radial | PSNR (dB): 35.50 ± 2.24 -> 36.23 ± 2.21 | Helped |
| Use neural operators instead of E2E-VN++ | MRI reconstruction on fastMRI knee · fastMRI knee, 4× Poisson | PSNR (dB): 33.01 ± 2.67 -> 33.42 ± 2.34 | Helped |
| Use neural operators instead of E2E-VN++ | MRI reconstruction on fastMRI knee · fastMRI knee, 4× Gaussian | PSNR (dB): 30.65 ± 2.55 -> 31.25 ± 2.70 | Helped |
| Use neural operators instead of E2E-VN | MRI reconstruction inference · MRI reconstruction, inference | Inference time (s): 0.104 -> 0.158 | Hurt |
| Use neural operators instead of diffusion reconstruction | MRI reconstruction inference · MRI reconstruction, inference | Inference time (s): 93.84 -> 0.158 | Helped |
Surprising findings
- The standard E2E-VN overfits to rectilinear patterns and drops sharply on irregular patterns (radial, Gaussian, Poisson), with PSNR falling to around 23 dB.
Limitations
- Only one neural operator design (DISCO) explored.
- Evaluates image reconstruction metrics only, without diagnostic accuracy assessment.
- Experiments limited to fastMRI knee and brain datasets.
Experimental setup
Datasets: fastMRI knee fastMRI brain
Benchmarks: fastMRI knee fastMRI brain
Metrics: PSNR SSIM NMSE inference time
Baselines: E2E-VN E2E-VN++ U-Net CSGM ScoreMRI PnP-DM l1-Wavelet
Builds on: End-to-End VarNet DISCO convolutions Fourier neural operator U-Net
Closest papers by meaning
- Training Deep Learning Based Image Denoisers From Undersampled Measurements Without Ground Truth and Without Image Prior (CVPR 2019 · similarity 0.90)
- MRI Reconstruction with Interpretable Pixel-Wise Operations Using Reinforcement Learning (AAAI 2020 · similarity 0.89)
- UGoDIT: Unsupervised Group Deep Image Prior Via Transferable Weights (NeurIPS 2025 · similarity 0.87)
- Efficient Noise Calculation in Deep Learning-based MRI Reconstructions (ICML 2025 · similarity 0.87)
- Progressive Divide-and-Conquer via Subsampling Decomposition for Accelerated MRI (CVPR 2024 · similarity 0.86)
- UFC-Net: Unrolling Fixed-point Continuous Network for Deep Compressive Sensing (CVPR 2024 · similarity 0.86)
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.