Hypernetwork Design for Self-Supervised Domain Adaptation in Compressive Sensing MRI Reconstruction

This paper presents a novel hypernetwork design for self-supervised domain adaptation in compressed sensing MRI (CS-MRI) reconstruction. Unlike conventional models, which often fail to generalize across distribution shifts in MRI data, our hypernetwork generates task-specific parameters for a primary reconstruction network based on learned regularization weights, dynamically adjusting to new test data. Leveraging test-time training (TTT), our hypernetwork significantly improves reconstruction accuracy on previously unseen data distributions. Evaluations using the fastMRI dataset demonstrate a notable gain in Peak Signal-to-Noise Ratio (PSNR) by 3.6 dB and Structural Similarity Index (SSIM) by 0.05 over baseline models, such as U-Net and VarNet, especially in out-of-distribution settings.