CMANet: A TCN-RMamba-Attention Network for Surgical Phase Online Recognition

Surgical phase recognition is a critical technology in assistive therapy, aiding physicians in enhancing surgical efficiency and postoperative assessment. Currently, deep learning approaches have been extensively applied to the task of surgical phase recognition. However, methods that rely solely on spatial features are prone to phase shaking. Notably, approaches based on Temporal Convolutional Network (TCN), Transformers or Mamba have demonstrated significant efficacy in handling temporal features. Nevertheless, existing methodologies face several challenges: 1) there is currently no method that integrates all three of the aforementioned approaches simultaneously to leverage their respective advantages; and 2) previous attention mechanisms primarily reduce computational costs through local attention, sparse attention, downsampling, which may inadvertently compromise the performance of attention. To address these gaps, we propose a TCN-Residual Mamba (RMamba)-Attention Network (CMANet), which comprises two key components: spatial feature extraction and prediction. Our main innovation lies in the prediction component, which is further divided into two prediction stages. Both prediction stages employ the same structure, organically integrating TCN, RMamba, and differential attention (DA) mechanisms. TCN and RMamba exhibit linear complexity in long sequence modeling and achieve better performance at a lower cost. In contrast, DA employs a differential mechanism to drastically minimize computational costs while demonstrating superior performance in long-sequence tasks requiring strong coping or contextual learning capabilities. The efficacy of the proposed method is validated on two benchmark datasets for surgical phase recognition, with experimental results demonstrating its effectiveness.