Research map / 3D vision / Depth and stereo
Progressive Divide-and-Conquer via Subsampling Decomposition for Accelerated MRI
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 Progressive Divide-and-Conquer (PDAC), which decomposes the severe k-space subsampling into moderate stages unfolded as a DUN with a learned degradation predictor and severity conditioning, evaluated on fastMRI and Stanford2D FSE.
Problem
Accelerated MRI reconstruction from highly undersampled k-space is highly ill-posed; conventional deep unfolding networks try to recover all missing information in the entire null space at every iteration, which leads to cumulative errors and subpar reconstructions.
Approach
Decompose the subsampling operator into a product of moderate degradations and progressively recover one segment of the null space per unfolding iteration, learning the decomposition (predicted sampling mask) as an auxiliary task with a degradation predictor and conditioning each stage on degradation severity via adaptive layer norm.
Key results
- On 8x multi-coil fastMRI knee, HUMUSNet+PDAC improves PSNR from 36.84 dB to 37.12 dB over the HUMUSNet backbone.
- On Stanford2D FSE 8x multi-coil, PDAC reaches 36.77 dB PSNR vs 35.43 dB for HUMUSNet (about 1.34 dB gain).
- On single-coil fastMRI 8x, HUMUSNet+PDAC reaches 31.16 dB PSNR vs 31.04 dB for HUMUSNet, with state-of-the-art PSNR/SSIM at 8x.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add PDAC to HUMUSNet | HUMUSNet · fastMRI knee, multi-coil 8× | PSNR: 36.84 -> 37.12 | Helped |
| add PDAC to HUMUSNet | HUMUSNet · Stanford2D FSE, multi-coil 8× | PSNR: 35.43 -> 36.77 | Helped |
| add PDAC to HUMUSNet | HUMUSNet · fastMRI knee, single-coil 8× | PSNR: 31.04 -> 31.16 | Helped |
| add adaptive degradation prediction (ablation) | PDAC · Stanford2D FSE | PSNR: 36.46 -> 36.77 | Helped |
| add decomposed degradation loss Lprob (ablation) | PDAC · Stanford2D FSE | PSNR: 36.48 -> 36.77 | Helped |
| add severity conditioning module (ablation) | PDAC · Stanford2D FSE | PSNR: 36.66 -> 36.77 | Helped |
| use fine-to-coarse rather than coarse-to-fine sampling budget schedule (ablation) | PDAC · Stanford2D FSE | PSNR: 36.77 -> 34.49 | Hurt |
| use uniform rather than coarse-to-fine sampling budget schedule (ablation) | PDAC · Stanford2D FSE | PSNR: 36.77 -> 36.66 | Hurt |
Surprising findings
- Removing the learned degradation prediction (using random decomposition masks with the same budget) costs about 0.31 dB PSNR, showing the learned decomposition matters.
Limitations
- Gains over the HUMUSNet backbone are modest on fastMRI knee (e.g., 0.28 dB multi-coil, 0.12 dB single-coil 8x).
- Single-slice input used for fair comparison, differing from original HUMUSNet setting.
- Cartesian mask assumption for the learned decomposition.
Experimental setup
Datasets: fastMRI knee Stanford2D FSE
Benchmarks: fastMRI knee Stanford2D FSE
Metrics: PSNR SSIM NMSE
Baselines: ESPIRiT U-Net E2E-VarNet HUMUSNet CS KIKI-Net Kiu-net SwinIR D5C5 OUCR ReconFormer
Builds on: HUMUSNet half-quadratic splitting ADMM-Net MoDL adaLN
Closest papers by meaning
- Breaking Information Isolation: Accelerating MRI via Inter-sequence Mapping and Progressive Masking (AAAI 2025 · similarity 0.90)
- Improving Deep Learning for Accelerated MRI With Data Filtering (NeurIPS 2025 · similarity 0.87)
- Efficient Noise Calculation in Deep Learning-based MRI Reconstructions (ICML 2025 · similarity 0.87)
- A Unified Model for Compressed Sensing MRI Across Undersampling Patterns (CVPR 2025 · similarity 0.86)
- Null Space Matters: Range-Null Decomposition for Consistent Multi-Contrast MRI Reconstruction (AAAI 2024 · similarity 0.86)
- D2SA: Dual-Stage Distribution and Slice Adaptation for Efficient Test-Time Adaptation in MRI Reconstruction (NeurIPS 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.