To explore how computer vision can extract actionable insights from surgical video, enhancing documentation and decision-making in surgery.
Approach:
Identifying Surgical Phases: Computer vision models were developed to identify individual phases of robotic prostatectomy with accuracy in the low- to mid-90% range.
Automated Operative Reports: AI-generated operative reports were created from surgical videos, achieving 87.3% accuracy compared to 72.8% for surgeon-written reports.
Evaluating Bladder Lesions: A computer vision model was trained to predict tumor histology from video during transurethral resection, achieving AUCs of 0.829 and 0.867 in different cohorts.
Automating Instrument Counting: A proof-of-concept project demonstrated a model's ability to detect and count surgical instruments in overlapping conditions.
Key Findings:
AI-generated operative reports showed higher accuracy than those written by surgeons.
The potential for video-linked operative reports could enhance surgical documentation.
Computer vision models can assist in evaluating ambiguous bladder lesions.
Automation of routine tasks like instrument counting could improve surgical efficiency.
Interpretation:
The findings suggest that computer vision has the potential to enhance surgical documentation, decision-making, and efficiency, although further validation is needed.
Limitations:
AI models may struggle with uncommon situations and subjective data can introduce noise.
Models can overfit to training data, necessitating validation across diverse settings.
Conclusion:
The research indicates that the applications of computer vision in surgery are just beginning to be explored.