Evaluating robotic-assisted surgery training videos with multi-task convolutional neural networks.

Area of Science:

• Surgical Education
• Artificial Intelligence in Medicine
• Robotics and Automation

Background:

• Assessing surgical proficiency is crucial for training effective surgeons.
• Objective evaluation of surgical skills, like the urethrovesical anastomosis, remains challenging.
• Current methods often rely on subjective human scoring.

Purpose of the Study:

• To investigate the efficacy of an automated algorithm in replacing human scoring of surgical trainees.
• To develop and evaluate a neural network model for predicting surgical proficiency scores (GEARS score) from video data.
• To assess the algorithm's ability to differentiate between various proficiency levels (novice to expert).

Main Methods:

• Utilized video recordings of surgeons performing urethrovesical anastomosis on synthetic tissue.
• Developed an algorithm to track surgical instrument locations and key point positions over time.
• Trained a multi-task convolutional neural network using positional features to infer GEARS score sub-categories.

Main Results:

• The automated system achieved scores matching manual inspection in 86.1% of all GEARS sub-categories.
• The model successfully differentiated between novice and expert proficiency levels in 83.3% of videos.
• The artificial neural network approach demonstrated feasibility for evaluating novice and intermediate surgeons.

Conclusions:

• Automated assessment of surgical proficiency using artificial intelligence is a viable approach.
• The developed algorithm shows significant potential for objective and consistent surgical skill evaluation.
• Further research is required to refine the system for accurate assessment of expert-level surgeons.