Surgical skills assessment from robot assisted surgery video data

For the sake of objectively assessing and saving time, demanding for surgical skill automatically assessment increased with each passing day. Prior works on this task were mainly kinematic data based, adopted descriptive statistics, hidden Markov models (HMMs) and descriptive curve coding (DCC) evaluation methods. As video based method has wider application prospect, we present a framework for automated assessment of the expertise level of surgeons using the global rating scores (GRS) criteria based on robot assisted surgery videos. We represent the motion dynamics via space temporal interest point (STIP) and improved dense trajectory (iDT) features. Bag-of-features (BoF) is used to derive a histogram to represent video followed by support vector machine (SVM) with a rbf kernel to classify surgical skill level. The framework is tested on robot assisted surgery videos of surgeons with different expertise levels performing basic surgical tasks. By using leave-one-super-trail-out (LOSO) method, we obtain the mean accuracy of 79.29% / 76.79%, 80.71% / 83.81% and 72.57% / 76.65% on the basis of STIP/iDT representation for suturing, knot tying and needle passing surgical tasks, respectively. Compared with kinematic data based results, this study clearly demonstrated the ability of video based assessment method to distinguish between novice and expert performance of robotic assisted surgery.