Towards an AI-based assessment model of surgical difficulty during early phase laparoscopic cholecystectomy

Abstract

ABSTRACT Laparoscopic cholecystectomy (LC) is the standard surgical treatment for patients with gallstone disease, ranging from symptomatic cholelithiasis to severe cholecystitis. As there is high variability in operative findings during LC, it is important to assess the surgical difficulty objectively. This is critical for further developing and using future artificial intelligence algorithms in LC surgery, as it allows more reliable benchmarking between surgeons and can help in surgical OR planning. In this study, deep learning models were trained to assess the level of difficulty in the first phase of the procedure. 93 LC videos recorded at the Meander Medical Center were included. A modified Nassar scale was used to annotate gallbladder difficulty (grade 1–3) and adhesion presence (grade 1–3). Various models, were trained on different label combinations;binary and multi-label. On the multi-label test set, the best model reached accuracies of 66% and 40% for the classification of the gallbladder and adhesions, respectively. The best binary models classifies gallbladder difficulty grade 3 vs 1,2 with an accuracy of 88%, and gives an accuracy of 82% for grade 1 vs 2,3. This work shows the potential of difficulty understanding in surgical scenery based on early phase endoscopic video.