Automated benchmarking of surgical skills using machine learning

Abstract

Surgical trainees are required to acquire specific skills during the course of their residency before performing real surgeries. Surgical training involves constant practice of skills and seeking feedback from supervising surgeons, who generally have a packed schedule. The process of manual assessment makes the whole training cycle extremely cumbersome and inefficient. Having automated assessment systems for surgical training can be of great value to medical schools and teaching hospitals. The aim of this PhD research is to develop machine learning based methods for assessment of surgical skills from basic tasks to complex robot-assisted procedures. Specifically, this thesis will cover details of (1) developing novel motion based features for basic surgical skills assessment in open and robotic surgical training, (2) developing unsupervised and supervised methods for recognizing individual steps of complex robot-assisted (RA) surgical procedures, (3) generating automated score reports for RA surgical procedures, and (4) producing video highlights to indicate which parts of the surgical task most effected the final surgical skill score. Positive results from experiments conducted confirms the feasibility of providing automated skill based feedback to surgeons.