Automated Surgical Phase Recognition in Unilateral Biportal Endoscopic Spine Surgery Using Deep Learning

Study Design: Retrospective observational study. Objective: To develop and evaluate an artificial intelligence (AI)–based model for automated surgical phase recognition in unilateral biportal endoscopic (UBE) spine surgery. Summary of Background Data: UBE spine surgery has gained widespread adoption, and understanding the surgical workflow has become essential for education and quality improvement. Meanwhile, AI has demonstrated strong performance in surgical phase recognition in laparoscopic and cataract surgeries. However, no previous study has applied this technology to spine surgery, where video analysis has been limited by technical and practical constraints. Methods: Thirty-six cases of single-level UBE decompression for lumbar spinal stenosis were retrospectively analyzed. Surgical videos were annotated into five phases. A 50-layer residual neural network (ResNet-50) was used for spatial feature extraction, and a multistage temporal convolutional network (MS-TCN) modeled temporal dependencies. Both non-causal (past and future frames) and causal (past frames only) versions were trained and validated using four-fold cross-validation and evaluated by accuracy and F1 score. In addition, phase-specific durations were compared between board-certified spine surgeons (BCS) and a non-board-certified spine surgeon (NBCS). Results: The non-causal and causal models achieved overall accuracies of 86.4% and 84.7%, and F1 scores of 0.753 and 0.727, respectively. Bone resection and yellow ligament resection phases showed high true-positive rates (>90%) in both models. Surgical time was significantly longer in the NBCS group across all major phases, with the largest difference observed in yellow ligament resection. AI-predicted and manually annotated phase durations showed consistent results. Conclusions: This study is the first to apply AI-based surgical phase recognition to spine surgery. Combining ResNet-50 and MS-TCN enabled accurate and consistent phase prediction in UBE decompression. The proposed system has potential to enhance surgical education, objective performance assessment, and real-time intraoperative support.