Automated recognition of surgical phases is a prerequisite for computer-assisted analysis of surgeries. The research on phase recognition has been mostly driven by publicly available datasets of laparoscopic cholecystectomy (Lap Chole) videos. Yet, videos observed in real-world settings might contain challenges, such as additional phases and longer videos, which may be missing in curated public datasets. In this work, we study (i) the possible data distribution discrepancy between videos observed in a given medical center and videos from existing public datasets, and (ii) the potential impact of this distribution difference on model development. To this end, we gathered a large, private dataset of 384 Lap Chole videos. Our dataset contained all videos, including emergency surgeries and teaching cases, recorded in a continuous time frame of five years. We observed strong differences between our dataset and the most commonly used public dataset for surgical phase recognition, Cholec80. For instance, our videos were much longer, included additional phases, and had more complex transitions between phases. We further trained and compared several state-of-the-art phase recognition models on our dataset. The models’ performances greatly varied across surgical phases and videos. In particular, our results highlighted the challenge of recognizing extremely under- represented phases (usually missing in public datasets); the major phases were recognized with at least 76 percent recall. Overall, our results highlighted the need to better understand the distribution of the video data phase that recognition models are trained on.
