Abstract
Ulcerative Colitis (UC) is an incurable inflammatory bowel disease that leadsto ulcers along the large intestine and rectum. The increase in the prevalenceof UC coupled with gastrointestinal physician shortages stresses the healthcaresystem and limits the care UC patients receive. A colonoscopy is performed todiagnose UC and assess its severity based on the Mayo Endoscopic Score (MES).The MES ranges between zero and three, wherein zero indicates no inflammationand three indicates that the inflammation is markedly high. ArtificialIntelligence (AI)-based neural network models, such as convolutional neuralnetworks (CNNs) are capable of analyzing colonoscopies to diagnose anddetermine the severity of UC by modeling colonoscopy analysis as a multi-classclassification problem. Prior research for AI-based UC diagnosis relies onsupervised learning approaches that require large annotated datasets to trainthe CNNs. However, creating such datasets necessitates that domain expertsinvest a significant amount of time, rendering the process expensive andchallenging. To address the challenge, this research employs self-supervisedlearning (SSL) frameworks that can efficiently train on unannotated datasets toanalyze colonoscopies and, aid in diagnosing UC and its severity. A comparativeanalysis with supervised learning models shows that SSL frameworks, such asSwAV and SparK outperform supervised learning models on the LIMUC dataset, thelargest publicly available annotated dataset of colonoscopy images for UC.