Robustness Against Out of Distribution Video Frames in Online Surgical Workflow Recognition with Temporal Convolutional Networks
Amirhossein Bayat, Kadir Kirtac, Salih karagoz, Julien Schwerin, Michael Stenzel, Marco Smit, Florian Aspart
The automatic recognition of surgical phase based on laparoscopic videos is a pre-requisite to diverse AI application on surgeries. Online surgical phase recognition is commonly achieved using two-stages models combining (i) a spatial feature extraction at the frame level with a (ii) temporal model. Yet, this online surgical phase recognition is a challenging task in real-world scenarios. For example, the camera might be temporally extracted of the body during surgeries (e.g., to be cleaned). The Out-of-body (OOB) phases have out-of-distribution spatial features and have unpredictable occurrence which affect the temporal model performance. We propose a simple, yet effective, mechanism to robustify our temporal model against OOB phases. Our solution leverages the two-stages structure of surgical phase model predictions. We train and test our model on a large scale real-world dataset of laparoscopic cholecystectomy videos and show the effectiveness of our approach.
Wednesday 6th July
Poster Session 1.2 - onsite 11:00 - 12:00, virtual 15:20 - 16:20 (UTC+2)