Research map / Architectures and efficiency / Attention and transformers
Breaking Information Isolation: Accelerating MRI via Inter-sequence Mapping and Progressive Masking
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 IMA, a deep unfolding network for multi-sequence MRI reconstruction that couples T2/PD sequences via explicit inter-sequence mapping, cross-sequence converging attention, and a dilation-contraction progressive masking scheme.
Problem
Accelerating multi-sequence MRI reconstruction with deep unfolding networks, where current methods learn sequences in isolation and keep the sampling mask fixed, causing information isolation and poor convergence.
Approach
Explicitly map the reference sequence to the target via two learnable auxiliary matrices (X = YA + ~YB), couple sequences with a cross-sequence converging attention, and update the k-space mask across unfolding stages via a dilation-contraction masking mechanism that moves information from known to unknown regions.
Key results
- On IXI 8x radial, IMA reaches 41.948 dB PSNR, ahead of the strongest baselines (MC-VarNet and RNU).
- On fastMRI 8x random, IMA reaches 28.764 dB PSNR and 0.813 SSIM, ranking first across all settings in Table 1.
- Ablation on IXI 8x radial: DCM adds about 2.13 dB PSNR and CSCA about 1.18 dB, with no added parameters for DCM.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add explicit inter-sequence mapping, CSCA, and DCM | IMA · IXI and fastMRI, random and radial masks at 4× and 8× | PSNR: 43.984 vs 42.323; 50.810 vs 48.374; 41.291 vs 39.894; 41.948 vs 40.901; 30.125 vs 29.949; 31.518 vs 31.228; 28.764 vs 28 | Helped |
| add CSCA (ablation) | IMA · IXI (8×) radial | PSNR: 40.767 -> 41.948 | Helped |
| add CSCA (ablation) | IMA · IXI (8×) radial | SSIM: 0.990 -> 0.992 | Helped |
| add DCM (ablation) | IMA · IXI (8×) radial | PSNR: 39.814 -> 41.948 | Helped |
| add DCM (ablation) | IMA · IXI (8×) radial | SSIM: 0.989 -> 0.992 | Helped |
| use DCM kernel 1 rather than kernel 2 (ablation) | IMA · IXI (8×) radial | PSNR: 40.286 -> 41.948 | Helped |
| use DCM kernel 1 rather than kernel 3 (ablation) | IMA · IXI (8×) radial | PSNR: 40.872 -> 41.948 | Helped |
| use DCM kernel 1 rather than kernel 2 (ablation) | IMA · fastMRI (8×) radial | PSNR: 28.167 -> 28.493 | Helped |
| use DCM kernel 1 rather than kernel 3 (ablation) | IMA · fastMRI (8×) radial | PSNR: 28.319 -> 28.493 | Helped |
| increase unfolding stages from 6 to 12 (ablation) | IMA · IXI and fastMRI (8× random mask) | PSNR | Helped |
Limitations
- Evaluated only on two small public datasets (IXI and knee fastMRI) with 1D Cartesian and 2D radial masks.
- DCM kernel and hyperparameters (delta, ratio, T) are chosen empirically per mask and sampling pattern.
Experimental setup
Datasets: IXI fastMRI knee
Benchmarks: IXI fastMRI knee
Metrics: PSNR SSIM P values Params GMACs
Baselines: Unet MICCAN DuDoRNet DuDoCAF MC-VarNet RNU
Builds on: deep unfolding networks half quadratic splitting RNU MC-VarNet
Closest papers by meaning
- Progressive Divide-and-Conquer via Subsampling Decomposition for Accelerated MRI (CVPR 2024 · similarity 0.90)
- Null Space Matters: Range-Null Decomposition for Consistent Multi-Contrast MRI Reconstruction (AAAI 2024 · similarity 0.88)
- Boosting ViT-based MRI Reconstruction from the Perspectives of Frequency Modulation, Spatial Purification, and Scale Diversification (AAAI 2025 · similarity 0.86)
- Improving Deep Learning for Accelerated MRI With Data Filtering (NeurIPS 2025 · similarity 0.85)
- Using Powerful Prior Knowledge of Diffusion Model in Deep Unfolding Networks for Image Compressive Sensing (CVPR 2025 · similarity 0.85)
- A Unified Model for Compressed Sensing MRI Across Undersampling Patterns (CVPR 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.