Development of an artificial intelligence algorithm for automated surgical gestures annotation.

Area of Science:

• Robotics in Surgery
• Artificial Intelligence in Medicine
• Surgical Skill Assessment

Background:

• Manual annotation of surgical gestures is time-consuming and subjective.
• Objective assessment of surgical quality is crucial for training and patient outcomes.
• Robot-assisted surgery provides a platform for detailed procedural analysis.

Purpose of the Study:

• To develop and evaluate a recurrent neural network (RNN) for automated surgical gesture annotation.
• To analyze gestures in simulated robot-assisted radical prostatectomies.
• To establish an efficient method for surgical quality assessment.

Main Methods:

• A dataset of 161 videos with five surgical gestures was manually annotated.
• A two-part neural network model was created: a Vision Transformer feature extractor and an LSTM-based classification head.
• The model was trained and validated, then tested on its ability to classify surgical gestures.

Main Results:

• The neural network achieved an Area Under the Curve (AUC) of 0.95 and an F1-score of 0.71.
• High accuracies (0.84-0.97) and specificities (0.90-0.99) were observed for gesture classification.
• Average Total Agreement (a measure of overlap) ranged from 0.72 to 0.91 across gesture classes.

Conclusions:

• A high-performing neural network for automated surgical gesture analysis was successfully developed.
• The model demonstrates potential for objective and efficient assessment of surgical procedures.
• Future work will focus on annotating real surgical videos and evaluating prediction of patient outcomes.