Position-Aware Singular Value Shrinkage for Unfolded Dynamic MRI Reconstruction

Hongtao Chen, Zhe Yang, Wenrui Li, Guanghui Cheng

Keywords: MRI Reconstruction, Singular Value Shrinkage

Abstract: Dynamic MRI reconstruction benefits from low-rank priors to exploit spatiotemporal redundancy. Recent deep unfolding networks (DUNs) often adopt Singular Value Thresholding (SVT) to apply low-rank constraints. However, most methods apply uniform or globally scaled thresholds, ignoring the unequal importance of singular values and the resolution-dependent nature of dynamic MR images. This leads to suboptimal shrinkage and poor generalization across anatomical variations. Existing adaptive shrinkage techniques in classical models are not trainable and incompatible with end-to-end learning. To address these challenges, we propose a Position-Aware Adaptive Singular-Value Shrinkage (PASS) module that learns to perform context-aware SVT using spectral positional encoding and a neural gating mechanism. This enables selective preservation of important components while suppressing noise and redundancy. We integrate PASS into a deep unfolding network based on low-rank plus sparse decomposition, and introduce a multi-resolution training strategy to improve the adaptivity of PASS across varying anatomical scales and acquisition settings. Experimental results on two dynamic cardiac MRI datasets demonstrate that our method achieves superior reconstruction quality and generalization compared to existing SVT-based baselines. Our code will be available after acceptance.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 9590