BiasRecon hits top PSNR in cross-anatomy, cross-center and cross-modality MRI reconstruction with fewer than 100 tunable parameters
Synopsis
The work proposes BiasRecon, a bias-calibrated adaptation framework grounded in a minimal-intervention principle, which alternates frequency-guided prior calibration (learnable scalars α and β modulating low- and high-frequency components at each U-Net skip layer, adding only 2L parameters), score-based denoising, and adaptive regularization that uses Stein's Unbiased Risk Estimator (SURE) to tune the regularization parameter γ; trained on FastMRI knee (973 volumes) and evaluated without retraining on FastMRI-Brain (cross-anatomy), Stanford Knee (cross-center) and CMRxRecon (combined shift) under 8× Gaussian 1D undersampling, it improves PSNR over DDS by +1.19 dB, +2.06 dB and +1.60 dB respectively, and is best in 10 of 12 sampling/acceleration settings.
Interpretation
Through t-SNE feature visualization and frequency decomposition, the authors find that deeper layers and low-frequency components are domain-specific, while shallow layers and high-frequency components remain domain-agnostic. Prior work often treats distribution shift as uniform degradation; this analysis separates what transfers from what does not along both layer and frequency axes. Based on t-SNE and frequency-decomposition visualizations in Fig. 1(b,c); qualitative, with no quantitative transferability metric reported.
BiasRecon formulates open-world adaptation as alternating optimization: P1 frequency-guided prior calibration, P2 score-based denoising, P3 data fidelity with a proximity term, using fewer than 100 tunable parameters overall. Unlike domain-adaptation methods that retrain or fine-tune many parameters, this framework adapts at test time with very few parameters, and δ is initialized to identity so the original network is preserved. The method introduces two scalars α and β per U-Net layer (2L parameters total) and builds a self-supervised loss LSSL from complementary Gaussian masks M=MΛ+MΓ, requiring no ground truth.
SURE with a Monte Carlo approximation adaptively estimates the regularization parameter γ without ground truth, combined with sliding-window early stopping. It turns γ from a training-set-tuned constant into a per-sample, progressively updated quantity that matches test-time noise and k-space sampling quality. Ablation under cross-center 4× Gaussian 1D shows +1.05 dB from RPA alone, +2.28 dB from FPC alone, and +2.94 dB with both.
Open-world reconstruction is validated on four datasets: +1.19 dB cross-anatomy, +2.06 dB cross-center, +1.60 dB combined shift relative to DDS, and best in 10 of 12 sampling/acceleration settings with second best in the other 2. It covers cross-anatomy, cross-center and cross-modality shifts simultaneously, with no retraining at test time. Each test set randomly selects 100 volumes with 3 slices each, images unified to 320×320, coil sensitivity maps estimated by ESPiRiT, metrics PSNR/SSIM/LPIPS.
Perspective
The result targets undersampled reconstruction in multi-coil parallel MRI, with FastMRI knee (973 volumes) as the training set and FastMRI-Brain, Stanford Knee and CMRxRecon as three shift types, under Gaussian 1D and Uniform 1D sampling at 4× and 8× acceleration with images unified to 320×320. It enables deployment without retraining under unseen centers, anatomies and protocols, benefiting reconstruction researchers and engineering teams that want to reuse a single pretrained model; the prerequisite is access to test-time k-space measurements and coil sensitivity estimation (ESPiRiT in this work).
The transferability conclusion comes from t-SNE and frequency-decomposition visualizations, which are qualitative, so it remains unclear whether the pattern holds for other network architectures or modalities. The exact count of fewer than 100 parameters depends on the U-Net layer count L, expressed as 2L, and L is not given numerically. The SURE estimate relies on a Monte Carlo approximation and a perturbation scale ϵ, whose variance effect is not explored in the text. In addition, all open-world tests use FastMRI knee as the sole training source, so whether cross-center and cross-modality conclusions change with a different training source remains to be seen.