Uncertainty-aware network for fine-grained and imbalanced reflux esophagitis grading.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Computer-aided diagnosis (CAD) aids endoscopists in image analysis, improving diagnostic accuracy.
• Existing CAD research for gastroesophageal reflux disease is limited, particularly for reflux esophagitis (RE).

Purpose of the Study:

• To develop and evaluate an advanced CAD system for fine-grained classification of reflux esophagitis (RE).
• To address challenges in RE diagnosis using endoscopic images, including class imbalance.

Main Methods:

• Proposed an uncertainty-aware network incorporating handcrafted features.
• Utilized representation and classifier decoupling with metric learning for improved classification.
• Employed test-time augmentation for enhanced interpretability and uncertainty estimation.

Main Results:

• The proposed CAD system achieved 90.2% accuracy and a 90.1% F1 score on a hospital-collected dataset of over 3000 images.
• The method demonstrated superior performance compared to previous approaches for RE classification.

Conclusions:

• The developed uncertainty-aware network effectively diagnoses reflux esophagitis (RE) with high accuracy.
• This CAD system shows promise for assisting endoscopists and improving timely treatment for RE.