Surgical Skill Assessment Automation Based on Sparse Optical Flow Data

Objective skill assessment based personal feedback is a vital part of surgical training. Automated assessment solutions aim to replace traditional manual (experts’ opinion-based) assessment techniques, that predominantly requires the most valuable time commitment from senior surgeons. Typically, either kinematic or visual input data can be employed to perform skill assessment. Minimally Invasive Surgery (MIS) benefits the patients by using smaller incisions than open surgery, resulting in less pain and quicker recovery, but increasing the difficulty of the surgical task manyfold. Robot-Assisted Minimally Invasive Surgery (RAMIS) offers higher precision during surgery, while also improving the ergonomics for the performing surgeons. Kinematic data have been proven to directly correlate with the expertise of surgeons performing RAMIS procedures, but for traditional MIS it is not readily available. Visual feature-based solutions are slowly catching up to the efficacy of kinematics-based solutions, but the best performing methods usually depend on 3D visual features, which require stereo cameras and calibration data, neither of which are available in MIS. This paper introduces a general 2D image-based solution that can enable the creation and application of surgical skill assessment solutions in any training environment. A well-established kinematics-based skill assessment benchmark’s feature extraction techniques have been repurposed to evaluate the accuracy that the generated data can produce. We reached individual accuracy up to 95.74% and mean accuracy – averaged over 5 cross-validation trials – up to 83.54%. Additional related resources such as the source codes, result and data files are publicly available on Github (https://github.com/abc-irobotics/visdatasurgicalskill).