Robust multi-label surgical tool classification in noisy endoscopic videos

Area of Science:

• Computer Science
• Medical Imaging
• Machine Learning

Background:

• Surgical data science is an emerging field leveraging machine learning (ML) for analyzing surgical procedure recordings.
• Existing ML models for surgical tasks are limited by a lack of well-annotated datasets and inaccurate labels.
• Robust models are needed for accurate surgical tool classification from endoscopic videos.

Purpose of the Study:

• To develop a systematic methodology for robust surgical tool classification using noisy endoscopic videos.
• To introduce an intelligent active learning strategy for efficient dataset curation and label correction.
• To implement a semi-supervised self-training framework for improved classification performance.

Main Methods:

• An active learning strategy utilizing collective intelligence for minimal dataset identification and expert-guided label correction.
• A student-teacher model-based self-training framework for semi-supervised classification of 14 surgical tools.
• Techniques including weighted data loaders and label smoothing to handle difficult samples and class imbalance.

Main Results:

• The proposed methodology achieved an average F1-score of 85.88% with class weights in the ensemble model-based self-training.
• An F1-score of 80.88% was achieved without class weights, demonstrating robustness with noisy labels.
• The approach significantly outperformed existing methods in surgical tool classification.

Conclusions:

• The developed methodology provides a robust solution for surgical tool classification from noisy endoscopic videos.
• The combination of active learning and self-training effectively addresses data scarcity and label noise challenges.
• This work advances the development of reliable ML models for surgical data science applications.