Introduction:

In recent years, the use of artificial intelligence (AI) for interpretation of medical imaging has gained prominence and continues to be studied extensively. However, the use of AI-enabled computer vision for interpretation of surgical video has been limited. Leveraging AI for automated annotation of surgical video has potential applications in surgical training, performance assessment, correlation of intraoperative events to postoperative outcomes, and may eventually enable AI-powered real-time intraoperative decision support. We aim to develop a computer vision algorithm for automated video annotation of surgical steps during endoscopic transurethral resection of bladder tumors (TURBT).

Methods:

Full-length surgical videos from TURBT were manually annotated by a team of medical image annotators supervised by a fellowship-trained urologic oncologist. Surgical videos were labeled with each of the following surgical steps: (1) primary endoscopic evaluation, including tumor evaluation and identification of any relevant anatomic landmarks, (2) bladder tumor resection, and (3) surface coagulation and hemostasis. Manually annotated surgical videos were then used to train a computer vision AI algorithm to autonomously identify each of these surgical steps on TURBT video. Performance accuracy of the AI algorithm was compared to human annotations as the gold-standard.

Results:

A total of 108 full-length TURBT videos were included, which were subdivided into training (n=62), internal validation (n=19), and test (n=27) cohorts. Overall accuracy for the complete AI model was 86.3%. Per-step accuracy for (1) primary evaluation, (2) tumor resection, and (3) hemostasis was 84.4%, 95.3%, and 65.0%, respectively.

Conclusion:

We present results of an AI-enabled computer vision algorithm for automated annotation of TURBT surgical video footage. Automated video annotation has potential applications in surgical training, surgeon self-assessment, performance review, quality improvement, and may facilitate future efforts to correlate intraoperative surgical events to postoperative outcomes.

Funding: N/A