Research map / Label-efficient and robust learning / Self-supervised and contrastive
AeSPa : Attention-guided Self-supervised Parallel Imaging for MRI Reconstruction
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 AeSPa, a zero-shot scan-specific self-supervised parallel imaging method that uses all acquired k-space for inputs and targets, stabilized by attention-guided k-space selection (AKSM), iteration-wise k-space masking (IKM), and a cross-channel sensitivity-map constraint.
Problem
Scan-specific zero-shot MRI reconstruction without fully-sampled reference data performs poorly at high acceleration because it splits k-space into disjoint sets and suffers from training instability and limited image representations.
Approach
Use all undersampled k-space points both as network input and as loss targets, with an attention (Mamba-based) selection of non-sampled k-space points for final inference, iteration-varying random masking to stabilize training, and a cross-channel constraint on sensitivity maps that works at high reduction factors with few ACS lines.
Key results
- On FastMRI knee with 1D Gaussian sampling at R=8, AeSPa reaches 32.57 dB PSNR vs 32.88 dB for SPICER at R=4 baseline comparison; at R=4 it reaches 35.99 dB PSNR and 0.899 SSIM, the best among compared methods.
- On FastMRI brain with 1D Gaussian R=4, AeSPa achieves 39.08 dB PSNR and 0.956 SSIM.
- Average reconstruction time is 12 minutes, versus 10.5 min for SPICER, 21.5 min for SSDU and 44 min for ZS-SSL.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use AeSPa instead of SPICER | FastMRI knee PD reconstruction · 1D Gaussian R=4 | PSNR: 32.92 ± 3.21 -> 35.99 ± 3.04 | Helped |
| Use AeSPa instead of SPICER | FastMRI brain T1-Post reconstruction · 1D Gaussian R=4 | PSNR: 35.42 ± 2.89 -> 39.08 ± 2.04 | Helped |
| Add cross-channel constraint to sensitivity map estimation (ablation) | AeSPa · FastMRI knee PD, 1D Gaussian R=8 | PSNR: 32.43 ± 3.08 -> 32.57 ± 3.24 | Helped |
| Add IKM (ablation) | AeSPa · FastMRI knee PD, 1D Gaussian R=8 | PSNR: 29.70 ± 2.53 -> 31.24 ± 2.26 | Helped |
| Add AKSM (ablation) | AeSPa · FastMRI knee PD, 1D Gaussian R=8 | PSNR: 29.01 ± 2.73 -> 29.70 ± 2.53 | Helped |
| Set csm/ccm update cycle to 1/1 iterations (ablation) | AeSPa · FastMRI knee PD, 1D Gaussian R=8 | PSNR: 32.57 ± 3.24 -> 23.37 ± 4.69 | Hurt |
| Set csm/ccm update cycle to 1/10 iterations (ablation) | AeSPa · FastMRI knee PD, 1D Gaussian R=8 | PSNR: 32.57 ± 3.24 -> 21.88 ± 4.79 | Hurt |
| Set csm/ccm update cycle to 10/10 iterations (ablation) | AeSPa · FastMRI knee PD, 1D Gaussian R=8 | PSNR: 32.57 ± 3.24 -> 31.78 ± 2.86 | Hurt |
| Set AKSM selection parameter K to 600 (93.75%) (ablation) | AeSPa · FastMRI knee PD, 1D Gaussian R=8 | PSNR: 32.57 ± 3.24 -> 32.55 ± 3.36 | No clear effect |
| Set AKSM selection parameter K to 200 (31.25%) (ablation) | AeSPa · FastMRI knee PD, 1D Gaussian R=8 | PSNR: 32.57 ± 3.24 -> 32.45 ± 2.61 | No clear effect |
Surprising findings
- On the brain dataset with 1D Uniform sampling at R=8, AeSPa (30.49 dB) is slightly below SSDU (30.48 dB) in PSNR tie range and below L1-SPIRiT SSIM in some settings, indicating non-uniform gains across patterns.
Limitations
- Fixed number of iterations (2,000) without adaptive stopping criteria may compromise optimal reconstruction quality.
- Evaluated on a small set (25 patients, 50 slices) and only two anatomies with two reduction factors.
Experimental setup
Datasets: FastMRI knee (PD) FastMRI brain (T1-Post)
Benchmarks: FastMRI knee FastMRI brain
Metrics: PSNR SSIM reconstruction time
Baselines: ZF GRAPPA TGV CS-Wavelet L1-SPIRiT SSDU ZS-SSL SPICER
Builds on: SSDU ZS-SSL SPICER Mamba Deep Image Prior-style scan-specific training
Closest papers by meaning
- Adaptive Sampling of k-Space in Magnetic Resonance for Rapid Pathology Prediction (ICML 2024 · similarity 0.86)
- Boosting ViT-based MRI Reconstruction from the Perspectives of Frequency Modulation, Spatial Purification, and Scale Diversification (AAAI 2025 · similarity 0.85)
- Null Space Matters: Range-Null Decomposition for Consistent Multi-Contrast MRI Reconstruction (AAAI 2024 · similarity 0.85)
- D2SA: Dual-Stage Distribution and Slice Adaptation for Efficient Test-Time Adaptation in MRI Reconstruction (NeurIPS 2025 · similarity 0.84)
- A Unified Model for Compressed Sensing MRI Across Undersampling Patterns (CVPR 2025 · similarity 0.84)
- Fast MRI for All: Bridging Access Gaps by Training without Raw Data (NeurIPS 2025 · similarity 0.84)
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.