DEEP REINFORCEMENT LEARNING FOR ROBOT-ASSISTED SURGICAL TRAINING

Abstract

Minimally invasive surgery (MIS) has the advantage of smaller incision sizes and faster recovery compared to open surgery. Dexterous and complex surgical tools are needed to overcome the limits to operational motion due to smaller incisions. There is also a need to have proficient operation skill in a narrow space. As a result, training time of trainee surgeons is extended. To shorten the learning curve of hand-eye coordination in the laparoscopic surgery, guidance from a Deep Reinforcement Learning (DRL) intelligent agent is proposed to support the trainee in the surgical training. In this thesis, a Cyber-Physical System (CPS) is designed and proposed with a DRL supported laparoscopic surgery training system. A DRL agent is trained to drive the surgical instrument to finish a designated laparoscopic training task, and is capable to show movement suggestions for the surgeons after the DRL training is completed.