Automated operative phase identification in peroral endoscopic myotomy.

Area of Science:

• Medical technology
• Surgical innovation
• Artificial intelligence in medicine

Background:

• Computer vision (CV) and artificial intelligence (AI) are transforming medical image analysis.
• CV has been applied to identify surgical phases in laparoscopic videos.
• This study explores applying CV techniques to endoscopic procedures, specifically peroral endoscopic myotomy (POEM).

Purpose of the Study:

• To develop and evaluate a CV model for identifying surgical phases in POEM procedures.
• To assess the accuracy of AI in classifying distinct stages of POEM.

Main Methods:

• A deep-learning model, Convolutional Neural Network (CNN) plus Long Short-Term Memory (LSTM), named POEMNet, was trained on 30 POEM videos.
• Surgeons provided ground truth labels for surgical phases: submucosal injection, mucosotomy, submucosal tunnel, myotomy, and mucosotomy closure.
• POEMNet was tested on 20 additional POEM videos, and its performance was compared against surgeon annotations.

Main Results:

• POEMNet achieved an overall phase identification accuracy of 87.6%.
• The model demonstrated strong performance with mean unweighted and prevalence-weighted F1 scores of 0.766 and 0.875, respectively.
• Accuracy was higher for longer surgical phases (88.3%) compared to shorter phases (<5 min, 70.6%).

Conclusions:

• A deep-learning CV approach is effective for identifying phases in endoscopic procedures like POEM.
• The POEMNet model shows potential for integration into intra-operative decision-support systems.
• Further development of AI in CV can aid in post-operative risk prediction for endoscopic surgeries.