Research map / Architectures and efficiency / Attention and transformers
HUMUS-Net: Hybrid Unrolled Multi-scale Network Architecture for Accelerated MRI Reconstruction
Zalan Fabian, Berk Tinaz, Mahdi Soltanolkotabi
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace single-scale Swin with multi-scale MUST feature extractor (ablation) | unrolled SwinIR (Un-SS) · Stanford 3D | SSIM: 0.9319 -> 0.9357 | Improved |
| increase patch size from 1x1 to 2x2 (with 3x embedding dim) (ablation) | Un-MS · Stanford 3D | SSIM: 0.9357 -> 0.9171 | Hurt |
| process low-resolution convolutional features in transformer (HUMUS-Block) instead of full-resolution (ablation) | Un-MS · Stanford 3D | SSIM: 0.9357 -> 0.9449 | Improved |
| unroll SwinIR with sensitivity map estimator, small size (Un-SS) (ablation) | SwinIR · Stanford 3D | SSIM: 0.9336 -> 0.9319 | Hurt |
| replace U-Net denoiser with HUMUS-Block at matched parameters (ablation) | unrolled MRI network · Stanford 3D | SSIM: 0.9348 -> 0.9378 | Improved |
| add adjacent slice reconstruction to E2E-VarNet (ablation) | E2E-VarNet · Stanford 3D x8 | SSIM: 0.9432 -> 0.9457 | Improved |
| HUMUS-Net without ASR vs E2E-VarNet with ASR (ablation) | HUMUS-Net · Stanford 3D x8 | SSIM: 0.9457 -> 0.9467 | Improved |
| HUMUS-Net replaces E2E-VarNet | fastMRI knee x8 test · fastMRI knee test | SSIM: 0.8900 -> 0.8936 | Improved |
| HUMUS-Net replaces E2E-VarNet | Stanford 2D validation · Stanford 2D | PSNR: 33.9 -> 33.7 | Hurt |
| increase number of unrolled cascades (ablation) | HUMUS-Net · Stanford 3D x8 | SSIM | Improved |
| add validation split to training data | HUMUS-Net · fastMRI knee test | SSIM: 0.8936 -> 0.8945 | Improved |
| scale HUMUS-Net to HUMUS-Net-L (embed dim 66 to 96) | HUMUS-Net · fastMRI knee test | SSIM: 0.8936 -> 0.8944 | Improved |
Closest papers by meaning
- A Unified Model for Compressed Sensing MRI Across Undersampling Patterns (CVPR 2025 · similarity 0.88)
- Progressive Divide-and-Conquer via Subsampling Decomposition for Accelerated MRI (CVPR 2024 · similarity 0.87)
- Boosting ViT-based MRI Reconstruction from the Perspectives of Frequency Modulation, Spatial Purification, and Scale Diversification (AAAI 2025 · similarity 0.87)
- Improving Deep Learning for Accelerated MRI With Data Filtering (NeurIPS 2025 · similarity 0.87)
- Breaking Information Isolation: Accelerating MRI via Inter-sequence Mapping and Progressive Masking (AAAI 2025 · similarity 0.87)
- Efficient Noise Calculation in Deep Learning-based MRI Reconstructions (ICML 2025 · similarity 0.87)
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.
