SE-OHFM: A surgical phase recognition network with SE attention module

Abstract

With the development of robot-assisted minimally invasive surgery, enhanced automatic context recognition of surgical procedures is becoming critical to improve surgeon performance and patient safety. Deep neural networks can be efficiency at identifying surgical phases and analyzing surgical procedures. However, hard-to-identify frames in surgical videos tend to reduce the accuracy of recognition. This research adds the SE attention mechanism to ResNeXt101 to extract image features and recognize surgical phases, while using the Online Hard Frame Mapper (OHFM) to assist in recognizing hard frames. The experiment results show that the network with added attention can effectively extract the features of laparoscopic images. The proposed method achieved an accuracy of 85.8% on the M2CAI16 workflow challenge dataset.