Background/Objectives: The global adoption of minimally invasive surgery has generated extensive video repositories, creating new opportunities for data-driven surgical education and quality assessment. Automated surgical phase recognition enables objective trainee evaluation, standardized competency assessment, and systematic procedural documentation. However, class imbalance in surgical workflows, where certain phases comprise 30–35% of frames while others represent only 5–10%, remains a significant challenge. This imbalance causes models to underperform on underrepresented yet clinically important phases. Methods: A retrospective analysis of laparoscopic cholecystectomy videos is performed with the implementation of a frame—based deep learning framework to develop and validate a surgical phase recognition pipeline based on ResNet-50 architecture with transfer learning. The model was designed to extract features from surgical video frames and classify them into seven distinct phases, without incorporating temporal context. We used the Cholec80 dataset and applied class balancing techniques to address inherent class imbalance. Results: The model achieved a mean balanced accuracy of 91.80% across five folds with consistent performance across all surgical phases. Per-phase F1-scores ranged from 0.89 to 0.95, demonstrating balanced classification without significant performance degradation on underrepresented phases. The confusion matrix revealed prediction errors primarily among adjacent or visually similar phases, reflecting the inherent ambiguity of surgical phase transitions. In practical terms, the model correctly identified the surgical phase in more than 9 out of 10 frames, enabling reliable automated segmentation of the operative workflow. Conclusions: This study demonstrates that artificial intelligence can reliably analyze surgical video data, achieving consistent and accurate phase recognition in laparoscopic cholecystectomy.
Raptis et al. (Tue,) studied this question.