TBCA-SlowFast: A Spatiotemporal Network for Endoscopic Image-Based Round Window Membrane Puncture Recognition

Transcanal endoscopic minimally invasive ear surgery features mild trauma, efficient operation and fewer complications compared with traditional otologic surgery. Nevertheless, the confined ear canal space prevents the integration of traditional force sensors, resulting in the absence of real-time perception during round window membrane (RWM) puncture. To tackle this problem, this paper proposes TBCA-SlowFast, an endoscopic image-based recognition network for RWM puncture state detection. Based on the SlowFast baseline, a novel Temporal Branch Coordinate Attention (TBCA) module is embedded to strengthen fine-grained spatiotemporal feature extraction and suppress complex surgical background noise. A simulated RWM puncture dataset is established for model training and evaluation. Experimental results demonstrate that TBCA-SlowFast reaches a recognition accuracy of 88.23%, outperforming the original SlowFast by 1.96%. The lightweight TBCA module only brings slight increases in parameters (0.7M) and computation (14.1M FLOPs), striking a favorable balance between accuracy and real-time capability. This method can provide stable intraoperative feedback, reduce needle over-insertion risks, and improve the overall safety and reliability of transcanal minimally invasive ear surgery.