Actor-Critic Reinforcement Learning for Automatic Left Atrial Appendage Segmentation

Left atrial appendage (LAA) is a major thrombus formation site, potentially responsible for atrial fibrillation (AF)associated stroke. In analyzing the risk factor of the AF-patients, diagnosing the LAA anatomy plays a significant role. Therefore, an automatic segmentation of the LAA can facilitate an accelerated AF diagnosis. It can also help physicians in preprocedural planning of LAA closure, which is an implant-based strategy to prevent thromboembolism in LAA. However, the high anatomic variation of the LAA, and leaking through the adjacent left superior pulmonary vein yield major challenges in LAA segmentation. With some prior works generally relying on a manual annotation of a bounding box, fully automated segmentation approach is rare to be found. In this paper, we propose a fully automatic LAA segmentation method powered by an actor-critic reinforcement learning agent where the agent proposes necessary segmentation seeds to perform a geodesic distance-based segmentation. The proposed method could resolve all the major challenges of LAA segmentation. To the best of our knowledge, this is the first automated method for LAA segmentation. Compared to the previous methods, it performs the segmentation with a significantly greater efficiency, taking only 7.6 seconds.