Research map / Architectures and efficiency / Attention and transformers
Boosting ViT-based MRI Reconstruction from the Perspectives of Frequency Modulation, Spatial Purification, and Scale Diversification
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 FPS-Former, a ViT-based MRI reconstruction network with frequency modulation attention (Laplacian/Gaussian frequency pyramid), spatial purification attention (hash-based token grouping), and a multi-scale feed-forward network, outperforming SoTA on CC359, fastMRI and SKM-TEA.
Problem
Accelerated MRI reconstruction from undersampled k-space is ill-posed; ViT-based reconstructors lose high-frequency detail, attend across content-unrelated tokens, and lack multi-scale modeling.
Approach
Combines a Laplacian-style frequency pyramid that re-weights attention across frequency levels, LSH-like hashing that groups similar tokens for within-group attention, and a feed-forward network with parallel 3x3 and 5x5 depthwise convolution paths, plus a hybrid-expert convolutional refinement block.
Key results
- On CC359 at 4x AF, FPS-Former reaches 34.38 dB PSNR and 0.9321 SSIM, versus 34.16 dB for ReconFormer and 32.78 dB for AST.
- On SKM-TEA, it reports 35.64 dB PSNR at 4x AF and 32.85 dB at 8x AF, exceeding AST by 0.47 dB and 0.24 dB.
- Ablation removing FMAM drops PSNR by 0.24 dB on CC359 at 4x AF; removing SPAM drops it by 0.55 dB.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add FMAM (ablation) | FPS-Former · CC359 AF=4 | PSNR: 34.14 to 34.38; 0.24 | Helped |
| add SPAM (ablation) | FPS-Former · CC359 AF=4 | PSNR: 33.83 to 34.38; 0.55 | Helped |
| use SDFN instead of a conventional feed-forward network (ablation) | FPS-Former · CC359 AF=4 | PSNR: 33.95 to 34.38 | Helped |
| add HEFR (ablation) | FPS-Former · CC359 AF=4 | PSNR: 33.98 to 34.38; 0.40 | Helped |
| use 3 frequency pyramid layers instead of 2 (ablation) | FMAM in FPS-Former · CC359 AF=4 | PSNR: 34.30 to 34.38 | Helped |
| use 3 frequency pyramid layers instead of 4 (ablation) | FMAM in FPS-Former · CC359 AF=4 | PSNR: 34.36 to 34.38 | Helped |
| replace global MSA with SPAM (ablation) | FPS-Former · CC359 AF=4 | PSNR: 33.83 to 34.38 | Helped |
| replace window-based MSA with SPAM (ablation) | FPS-Former · CC359 AF=4 | PSNR: 33.89 to 34.38 | Helped |
| use SDFN instead of GDFN (ablation) | FPS-Former · CC359 AF=4 | PSNR: 33.82 to 34.38; 0.56 dB | Helped |
Limitations
- Gains over the strongest transformer baseline are modest on some settings (e.g., 0.25 dB over AST on fastMRI at 4x AF, 0.22 dB over ReconFormer on CC359 at 4x AF).
- Hyper-parameter analysis is deferred to supplementary material and not included in the text provided.
Experimental setup
Datasets: CC359 fastMRI SKM-TEA
Benchmarks: CC359 fastMRI SKM-TEA
Metrics: NMSE SSIM PSNR
Baselines: CS KIKI-Net UNet-32 D5C5 DCRCN VIT_Base SwinMR ReconFormer Restormer AST
Builds on: Restormer Vision Transformer Laplacian pyramid
Closest papers by meaning
- Null Space Matters: Range-Null Decomposition for Consistent Multi-Contrast MRI Reconstruction (AAAI 2024 · similarity 0.86)
- Breaking Information Isolation: Accelerating MRI via Inter-sequence Mapping and Progressive Masking (AAAI 2025 · similarity 0.86)
- Devignet: High-Resolution Vignetting Removal via a Dual Aggregated Fusion Transformer with Adaptive Channel Expansion (AAAI 2024 · similarity 0.85)
- Stochastic Deep Restoration Priors for Imaging Inverse Problems (ICML 2025 · similarity 0.85)
- Progressive Divide-and-Conquer via Subsampling Decomposition for Accelerated MRI (CVPR 2024 · similarity 0.85)
- Motion Artifact Removal in Pixel-Frequency Domain via Alternate Masks and Diffusion Model (AAAI 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.