用手术记忆增强时序建模，提升长视频手术阶段识别准确率

Surgical phase recognition from video enables various downstream applications. Transformer-based sliding window approaches have set the state-of-the-art by capturing rich spatial-temporal features. However, while transformers can theoretically handle arbitrary-length sequences, in practice they are limited by memory and compute constraints, resulting in fixed context windows that struggle with maintaining temporal consistency across lengthy surgical procedures. This often leads to fragmented predictions and limited procedure-level understanding. To address these challenges, we propose Memory of Surgery (MoS), a framework that enriches temporal modeling by incorporating both semantic interpretable long-term surgical history and short-term impressions. MoSFormer, our enhanced transformer architecture, integrates MoS using a carefully designed encoding and fusion mechanism. We further introduce step filtering to refine history representation and develop a memory caching pipeline to improve training and inference stability, mitigating shortcut learning and overfitting. MoSFormer demonstrates state-of-the-art performance on multiple benchmarks. On the Challenging BernBypass70 benchmark, it attains 88.0 video-level accuracy and phase-level metrics of 70.7 precision, 68.7 recall, and 66.3 F1 score, outperforming its baseline with 2.1 video-level accuracy and phase-level metrics of 4.6 precision, 3.6 recall, and 3.8 F1 score. Further studies confirms the individual and combined benefits of long-term and short-term memory components through ablation and counterfactual inference. Qualitative results shows improved temporal consistency. The augmented temporal context enables procedure-level understanding, paving the way for more comprehensive surgical video analysis.