SWNet: Surgical Workflow Recognition with Deep Convolutional Network

Surgical workflow recognition has been playing an essential role in computer-assisted interventional systems for modern operating rooms. In this paper, we present a computer vision-based method named SWNet that focuses on utilizing spatial information and temporal information from the surgical video to achieve surgical workflow recognition. As the first step, we utilize Interaction-Preserved Channel-Separated Convolutional Network (IP-CSN) to extract features that contain spatial information and local temporal information from the surgical video through segments. Secondly, we train a Multi-Stage Temporal Convolutional Network (MS-TCN) with those extracted features to capture global temporal information from the full surgical video. Finally, by utilizing Prior Knowledge Noise Filtering (PKNF), prediction noise from the output of MS-TCN is filtered. We evaluate SWNet for Sleeve Gastrectomy surgical workflow recognition. SWNet achieves 90% frame-level accuracy and reaches a weighted Jaccard Score of 0.8256. This demonstrates that SWNet has considerable potential to solve the surgical workflow recognition problem.