Self-supervised learning framework for efficient classification of endoscopic images using pretext tasks.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Accurate identification of anatomical landmarks in endoscopic video is crucial for diagnosing gastrointestinal diseases.
• Challenges include visual variability and limited annotated data, hindering traditional machine learning approaches.

Purpose of the Study:

• To develop a novel self-supervised learning (SSL) framework for enhanced feature learning from unlabeled endoscopic images.
• To improve the classification accuracy of key gastrointestinal anatomical regions (Z-line, esophagus, antrum/pylorus).
• To enhance model interpretability through attention mechanisms, transformers, and Grad-CAM visualization.

Main Methods:

• A self-supervised learning (SSL) framework integrating colorization, jigsaw puzzle solving, and patch prediction pretext tasks.
• Incorporation of attention mechanisms and transformer-based architectures for robust feature extraction.
• Utilized Grad-CAM for visualizing critical regions influencing model decisions.

Main Results:

• Achieved 98% classification accuracy for Z-line, esophageal, and antrum/pylorus regions.
• Demonstrated substantial improvements in precision, recall, and F1-score compared to conventional models.
• Validated effectiveness through ROC curves and confusion matrices, highlighting high performance across all classes.

Conclusions:

• The proposed SSL framework effectively enhances feature learning and classification accuracy for endoscopic anatomical landmark identification.
• Integration of attention, transformers, and Grad-CAM provides a scalable, interpretable, and clinically applicable solution.
• This methodology reduces reliance on extensive annotated datasets, advancing medical image analysis.