Abstract
OBJECTIVE: Accurate identification of surgical phases during cataract surgery is essential for improving surgical feedback and performance analysis. Time spent in each surgical phase is an indicator of performance, and segmenting out specific phases for further analysis can simplify providing both qualitative and quantitative feedback on surgical maneuvers. STUDY DESIGN: Retrospective surgical video analysis. SUBJECTS: One hundred ninety cataract surgical videos from the BigCat dataset (comprising nearly 4 million frames, each labeled with 1 of 11 nonoverlapping surgical phases). METHODS: Four machine learning architectures were developed for segmentation of surgical phases. Models were trained using cataract surgical videos from the BigCat dataset. MAIN OUTCOME MEASURES: Models were evaluated using metrics applied to frame-by-frame output and, uniquely in this work, metrics applied to phase output. RESULTS: The final model, CatStep, a combination of a temporally sensitive model (Inflated 3D Densenet) and a spatially sensitive model (Densenet169), achieved an F1-score of 0.91 and area under the receiver operating characteristic curve of 0.95. Phase-level metrics showed considerable boundary segmentation performance with a median absolute error of phase start and end time of just 0.3 seconds and 0.1 seconds, respectively, a segmental F1-score @70 of 0.94, an oversegmentation score of 0.89, and a segmental edit score of 0.92. CONCLUSION: This study demonstrates the feasibility of high-performance automated surgical phase identification for cataract surgery and highlights the potential for improved surgical feedback and performance analysis. FINANCIAL DISCLOSURES: Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.