Research map / Label-efficient and robust learning / Test-time adaptation
D2SA: Dual-Stage Distribution and Slice Adaptation for Efficient Test-Time Adaptation in 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 D2SA, a two-stage test-time adaptation for MRI reconstruction that first adapts patient-wise via an MR-INR (SIREN with mean/variance affine modulation) and then refines single slices with frozen CNN layers plus a learnable anisotropic diffusion module.
Problem
MRI reconstruction degrades under scanner, anatomy, dataset, modality and acceleration/sampling distribution shifts; existing TTA methods are slow (repeated single-slice training) or over-smooth with batch-wise self-supervision in data-scarce settings.
Approach
Model each patient's slices with an implicit neural representation whose features are modulated by learned mean (beta) and variance (alpha) shifts, then refine each slice with a DIP-style frozen-CNN network augmented by a learnable anisotropic diffusion module built from difference convolutions to preserve edges and cut adaptation time.
Key results
- On VarNet acceleration shift, FINE+MR-INR raises SSIM from 0.696 to 0.791 and PSNR from 21.39 to 25.30 over FINE.
- Anatomy shift (VarNet): SSDU+MR-INR+SST reaches 0.882 SSIM / 27.66 PSNR in 18.5 min/patient vs 52.5 min for DIP-TTT (0.878 / 27.67 / 0.312 ... reported 27.67 PSNR for DIP-TTT).
- U-Net anatomy ablation: MR-INR+SST with frozen CNN and trainable AD reaches 27.71 PSNR in 12.1 min/patient, the best among variants.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add MR-INR patient-wise adaptation and SST slice refinement to NR2N | VarNet · Dataset Shift | PSNR: 26.59 -> 28.76 | Helped |
| Add MR-INR patient-wise adaptation and SST slice refinement to NR2N | VarNet · Acceleration Shift | PSNR: 20.97 -> 28.89 | Helped |
| Add MR-INR patient-wise adaptation and SST slice refinement to NR2N | VarNet · Sampling Shift | PSNR: 21.15 -> 28.40 | Helped |
| Add MR-INR patient-wise adaptation to FINE (ablation) | U-Net · Anatomy Shift | PSNR: 25.98 -> 26.37 | Helped |
| Use trainable rather than frozen MR-INR latent codes (ablation) | FINE+MR-INR · Anatomy Shift | PSNR: 26.37 -> 26.48 | Helped |
| Add MR-INR to SST with trainable CNN (ablation) | U-Net · Anatomy Shift | PSNR: 27.22 -> 27.65 | Helped |
| Add trainable AD to MR-INR+SST with trainable CNN (ablation) | U-Net · Anatomy Shift | PSNR: 27.65 -> 27.54 | Hurt |
| Freeze MR-INR while using frozen CNN and trainable AD (ablation) | U-Net MR-INR+SST · Anatomy Shift | PSNR: 27.71 -> 27.41 | Hurt |
| Add trainable AD after freezing CNN layers (ablation) | U-Net MR-INR+SST · Anatomy Shift | PSNR: 27.38 -> 27.71 | Helped |
Surprising findings
- AD brings limited improvement when the CNN is fully trainable but yields notable gains when CNN layers are frozen.
- ZS-SSL does not generalise to end-to-end U-Net architectures in the authors' preliminary tests.
Limitations
- Evaluated only on public datasets; no real-world clinical undersampled MRI validation.
- Operates on 2D slices independently without exploiting full 3D correlations.
- Adaptation is offline per patient; no online or incremental learning for continuous streams.
Experimental setup
Datasets: fastMRI knee fastMRI brain fastMRI AXT2 fastMRI AXT1PRE Stanford knee
Benchmarks: fastMRI Stanford MRI dataset
Metrics: SSIM PSNR LPIPS adaptation time (mins/patient)
Baselines: DIP-TTT FINE Noiser2Noise (NR2N) SSDU ZS-SSL Zero-filling Non-TTA
Builds on: Functa SIREN Deep Image Prior (DIP-TTT) Noiser2Noise SSDU FINE
Closest papers by meaning
- Fully Convolutional Slice-to-Volume Reconstruction for Single-Stack MRI (CVPR 2024 · similarity 0.86)
- Progressive Divide-and-Conquer via Subsampling Decomposition for Accelerated MRI (CVPR 2024 · similarity 0.85)
- Boosting ViT-based MRI Reconstruction from the Perspectives of Frequency Modulation, Spatial Purification, and Scale Diversification (AAAI 2025 · similarity 0.85)
- Autoregressive Image Diffusion: Generation of Image Sequence and Application in MRI (NeurIPS 2024 · similarity 0.85)
- Self-Supervised Diffusion MRI Denoising via Iterative and Stable Refinement (ICLR 2025 · similarity 0.84)
- AeSPa : Attention-guided Self-supervised Parallel Imaging for MRI Reconstruction (CVPR 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.