Adverse Event Detection in Laparoscopic Videos

Human errors are one of the main causes of surgical adverse events. These events are reported by physicians which is an error-prone process. An automatic surgical adverse event detection system can reduce biases in the reports, provide quick feedback to surgeons, and help guide postoperative care. Few studies have developed such a system for laparoscopic procedures -- a major type of minimally-invasive procedure that has 15 million cases each year. We developed a deep learning system that can automatically detect two major surgical adverse events: bleeding and thermal injury. The system consists of a module that reduces camera motion in a scene, a module that extracts spatial-temporal features from videos, and a final module that takes into account the full temporal resolution of a procedure and makes predictions. This system significantly outperforms the prior state-of-the-art architectures developed for action recognition and adverse event detection in laparoscopic videos.M.A.S.2021-11-27 00:00:0