OSFENet: Object Spatiotemporal Feature Enhanced Network for Surgical Phase Recognition

Accurate recognition of surgical phases is paramount in open operating rooms, presenting an equally formidable challenge in developing a scene-aware system to achieve this task. However, previous studies have mainly focused on temporal information between video frames and consecutive frames, ignoring fine-grained details within individual frames, such as spatial information of surgical tools as well as motion information. In this paper, we propose a method to align temporal surgical tools and enhance video frame feature by integrating spatial features of surgical tools with temporal features, denoted as OSFENet. To validate the effectiveness of our proposed method for surgical phase recognition, we conducted experiments on three publicly available laparoscopic surgery datasets—Cholec80, M2cai16 and Autolaparo. Our experiments have yielded better results, achieving accuracy rates of \(91.8\, \pm \,3.9\%\) , \(90.1\, \pm \,6.1\%\) and \(\text{80.7\%}\) on Cholec80, M2cai16 and Autolaparo, respectively. This highlights that our use of detailed surgical tool information significantly improves the accuracy of surgical phase recognition.