Segmentation-aware MRI subsampling for efficient cardiac MRI reconstruction with reinforcement learning

Magnetic Resonance Imaging (MRI) scans, though highly detailed and non-invasive, take significantly longer than Computed Tomography (CT) scans and are sensitive to motion during acquisition. Accelerating MRI sampling and improving image reconstruction quality are crucial, especially for dynamic regions like the heart. Existing methods primarily enhance overall image quality but seldom target specific anatomic regions. In this paper, we propose a novel approach that combines segmentation and reinforcement learning to accelerate cardiac MRI sampling and enhance the reconstruction quality of cardiac regions. We design a policy network using reinforcement learning, where the input is a combination of the reconstructed image and the segmented category probability feature map, and the output determines the next k-space line to sample in a Cartesian setup. Retrospective testing on the ACDC (Automated Cardiac Diagnosis Challenge) cardiac segmentation dataset shows that our method significantly improves both the cardiac region and overall image quality compared to variants without segmentation rewards. Our approach ensures dynamically accelerated k-space sampling and surpasses current state-of-the-art reinforcement learning methods in producing diagnostically superior reconstructed cardiac MR images.