Deep Reinforcement Learning for Localization of the Aortic Annulus in Patients with Aortic Dissection

Accurate localization of the aortic annulus is key to several imaging tasks, like cross-sectional aortic valve plane estimation, aortic root segmentation, and annulus diameter measurements. In this project, we propose an end-to-end trainable deep reinforcement learning (DRL) algorithm aimed at identification of the aortic annulus in patients with aortic dissection. We trained 5 different agents on a dataset of 75 CT scans from 66 patients following a sequential model-upgrading strategy. We evaluated the effect of performing different image preprocessing steps, adding batch normalization and regularization layers, and changing terminal state definition. At each step of this sequential process, the model performance has been evaluated on a validation sample composed of 24 CTA scans from 24 independent patients. Localization accuracy was defined as the Euclidean distance between estimated and target aortic annulus locations. Best model results show a median localization error equal to 2.98 mm with an interquartile range equal to [2.25, 3.81] mm, and a failure rate (i.e., percentage of samples with localization error in validation data. We proved the feasibility of DRL application for aortic annulus localization in CTA images of patients with aortic dissection, which are characterized by a large variability in aortic morphology and image quality. Nevertheless, further improvements are needed to reach expert-human level performance.