Cross-Domain MRI Reconstruction using Transfer-Guided Untrained Networks and Reinforcement-Learned Adaptive Sampling

In this work, we propose a novel MRI reconstruction framework that integrates the benefits of transfer learning with adaptive sampling guided by reinforcement learning. While recent efforts in untrained deep neural networks have shown promising results for compressed sensing MRI (CS-MRI), their performance across different anatomical regions and undersampling patterns remains underexplored. To address this, we introduce a Transfer-Guided Untrained Network (TUNet) which utilizes shallow prior features from pretrained image models to initialize a variational untrained decoder. This allows for faster convergence and improved generalization across domains such as brain and knee MRI. Simultaneously, we propose a Sampling Policy Agent, trained via deep reinforcement learning, which dynamically adapts the k-space sampling trajectory to maximize reconstruction fidelity. Our hybrid pipeline is evaluated on the fastMRI dataset and demonstrates superior performance in PSNR, SSIM, and robustness to domain shifts, compared to traditional fixed-pattern sampling and purely untrained methods. The combined strategy of transfer-initialized untrained priors and learned sampling patterns opens new avenues for clinically robust, training-free MRI reconstruction.