Gastro

[Key Concept]

The rise in colorectal cancer underscores the need for early detection and accurate diagnosis of colorectal polyps. This dissertation introduces a novel method to enhance optical diagnosis by integrating Out-of-Distribution (OOD) detection and uncertainty quantification, assisting endoscopists in identifying hyperplastic (HP) and adenomatous (AD) polyps, and managing OOD polyps.Despite advancements, classifying polyps during colonoscopy is challenging due to variability among endoscopists and complex polyp types. Current Computer-Aided Diagnosis (CAD) systems achieve expert-level accuracy in distinguishing HP and AD polyps but lack interpretability and the ability to handle OOD polyps. This research develops a CAD system, Colood, that accurately classifies polyps, detects OOD polyps, and quantifies prediction uncertainty. Using advanced techniques for data augmentation, model training, and confidence calibration, Colood is tested on diverse datasets, including real colonoscopy videos, to improve colorectal polyp diagnosis and patient outcomes. The complete pipeline of Colood has achieved dominant performance over other techniques