Temporal segmentation of laparoscopic videos into surgical phases

Videos of laparoscopic surgeries need to be segmented temporally into phases so that surgeons can use the recordings efficiently in their everyday work. In this paper we investigate the performance of an automatic phase segmentation method based on instrument detection and recognition. Contrary to known methods that dynamically align phases to an annotated dataset, our method is not limited to standardized or unvarying endoscopic procedures. Phases of laparoscopic procedures show a high correlation to the presence of one or a group of certain instruments. Therefore, the first step of our procedure is the definition of a set of rules that describe these correlations. The next step is the spatial detection of instruments using a color-based segmentation method and a rule-based interpretation of image moments for the refinement of the detections. Finally, the detected regions are recognized with SVM classifiers and ORB features. The evaluation shows that the proposed technique find phases in laparoscopic videos of cholecystectomies reliably.