Quality-dependent Deep Learning for Safe Autonomous Guidewire Navigation

Cardiovascular diseases are the main cause ofdeath worldwide. State-of-the-art treatment often includes theprocess of navigating endovascular instruments through thevasculature. Automation of the procedure receives much at-tention lately and may increase treatment quality and unburdenclinicians. However, in order to ensure the patient’s safety theendovascular device needs to be steered carefully through thebody. In this work, we present a collection of medical criteriathat are considered by physicians during an intervention andthat can be evaluated automatically enabling a highly objectiveassessment. Additionally, we trained an autonomous controllerwith deep reinforcement learning to gently navigate within a2D simulation of an aortic arch. Among others, the controllerreduced the maximum and mean contact force applied to thevessel walls by 43% and 29%, respectively.