Surgical Video Phase Recognition Model Integrating Phase Localization Module and Temporal Context Method

Surgical phase recognition plays a crucial role in computer - assisted surgery systems. Its main purpose is to identify which stage of the surgery the video frames represent. In recent years, remarkable progress has been made in identifying surgical phases using deep - learning methods. However, existing methods mainly focus on improving the accuracy and precision of phase recognition, often neglecting the semantic consistency and temporal continuity of video frames within a phase. This easily leads to prediction errors for blurred and similar frames within a phase, posing challenges to reliable phase recognition. To overcome this problem, we propose a novel method for surgical video phase recognition and construct a temporal model combined with surgical phase localization. By predicting the start and end time points of surgical phases, the model improves the prediction consistency within surgical phases. The model consists of a two - stage training process. In the first stage, the spatial features of the video frame sequence are extracted. In the second stage, a temporal network is first used to perform frame-by-frame surgical phase prediction and surgical phase localization respectively. Finally, the prediction results of frame-by-frame surgical phases and the results of phase localization are fused to generate the final prediction information. This method effectively addresses the problem of intra - segment inconsistency in prediction results and significantly improves the accuracy and reliability of surgical phase recognition.