192 Automated Craniotomy Phase Detection Using Multi-Stage Temporal Convolutional Neural Networks

INTRODUCTION: Automated surgical phase recognition is essential to developing AI-driven models that improve intraoperative feedback and operating room efficiency. Most automated phase-detection methods target laparoscopic procedures. Here, we introduce a novel application for it in point-of-view craniotomy videos. METHODS: The TeCNO model, a causal, dilated two-stage TCN, classified craniotomy video segments into six phases: incision, myocutaneous flap reflection, burr hole placement, bone flap elevation, lateral sphenoid wing drilling, and dural opening. The dataset comprised 42 high-resolution, point-of-view craniotomy videos recorded at 25 frames per second. Each frame was manually labeled to one of six surgical phases. We used a 60-10-30 train/validation/test split. The model was trained and tested five times with Adam optimization, early stopping, weighted loss function, weighted sampler, and optimal F1 thresholding. RESULTS: The TeCNO model achieved overall average accuracy, precision, and recall of 0.71, 0.71, and 0.64, respectively. Specific average accuracies for phases 1-6 were 0.82, 0.66, 0.86, 0.49, 0.1, and 0.93, respectively. One-versus-all predictions using optimal F1 thresholding increased average accuracies to 0.96, 0.92, 0.90, 0.92, 0.82, and 0.97 and yielded per-phase average F1 scores of 0.81, 0.78, 0.80, 0.56, 0.50, and 0.77, respectively. There was high overall and per-phase average accuracy for all phases except five (lateral sphenoid wing drilling), which was frequently mistaken for six (dural opening). Phase five was not performed in every craniotomy and occasionally overlapped with actions performed in phase six. CONCLUSIONS: After optimal F1 thresholding, the MS-TCN model effectively and accurately detects craniotomy phases. Next steps include dataset enlargement and craniotomy completion time prediction based on individual surgeon characteristics.