COGNITIVE ENGINE AND DEEP REINFORCEMENT LEARNING FOR ROBOT-ASSISTED SURGERY

Abstract

Minimally invasive surgery (MIS) consists of surgical techniques which can dramatically reduce the trauma of the patients and subsequently decrease the risk of intraoperative blood loss, limit the risk of postoperative infection and improve postoperative recovery.

In this thesis, we researched and developed a cognitive engine for semantic-level surgical process management to provide patient-speci fic surgical procedures reasoning and storage in the knowledge database for re-querying. The proposed cognitive engine can perform semantic information recording, property-specific reasoning, knowledge storing, and knowledge sharing. To overcome the risk issues caused by tool-tissue interaction, we investigated the application of deep reinforcement learning (DRL) in surgical needle path planning. Two new path planning algorithms are proposed in this thesis to steer the flexible needle insertion in liver tumor radiofrequency ablation (RFA) surgery. Based on experimental results, both algorithms can improve the accuracy and robustness compared with that of the state-of-arts.