Automatic task recognition in a flexible endoscopy benchtop trainer with semi-supervised learning.

Area of Science:

• Medical Training and Simulation
• Computer Vision
• Machine Learning

Background:

• Inexpensive benchtop trainers are crucial for flexible endoscopy education.
• Current scoring systems rely on human raters, limiting scalability and cost-effectiveness.
• Automated rating using machine learning is needed to meet training demands.

Purpose of the Study:

• To develop and validate a machine learning approach for automating the rating of flexible endoscopy training.
• To enable automatic computation of task duration from endoscopic videos.
• To improve the accuracy of task recognition using a novel semi-supervised learning method.

Main Methods:

• A general and robust approach for recognizing training tasks from endoscopic videos was developed.
• State-of-the-art Convolutional Neural Network (CNN)-based methods were enhanced using a semi-supervised learning strategy.
• The approach utilized both labeled and unlabeled videos, assuming known task execution order for the latter.

Main Results:

• Two video datasets were used: 19 videos in examination conditions and 17 hours in self-assessment conditions.
• Mean task duration estimation error was 3.65s for the first dataset and 3.67s for the second.
• The semi-supervised learning approach reduced the average error from 5.63% to 3.67%.

Conclusions:

• This study represents a significant advancement in automating the assessment of flexible endoscopy trainees.
• The semi-supervised learning approach allows for easy scaling to large unlabeled training datasets.
• The developed methodology is applicable to other phase recognition tasks in medical training.