Additive Angular Margin Loss and Model Scaling Network for Optimised Colitis Scoring

Inflammatory bowel disease (IBD), in particular ulcerative colitis (UC), is required to be graded by endoscopists and is important for therapy monitoring. However, the accuracy of current endoscopic characterisation is operator dependant and can cause heterogeneous scoring leading to undesirable clinical outcomes for patients with IBD. Deep learning classification model for UC grading can be extremely helpful to clinicians allowing them for a comprehensive disease risk stratification. A systematic and continuous scoring is needed for better IBD patient stratification. While most methods in literature present a binary UC scoring, we propose a 3-way Mayo endoscopic subscore (MES) classification. In this context, we use an additive angular margin loss function in addition to cross-entropy loss to improve our model accuracy. We have evaluated multiple classification models and we demonstrate that the use of model scaling network together with added angular margin loss can provide improved results (over 10% improvement compared to ResNet-152).