CLoE: Curriculum Learning on Endoscopic Images for Robust MES Classification

Estimating disease severity from endoscopic images is essential in the assessment of ulcerative colitis, where the Mayo Endoscopic Subscore (MES) is widely used to grade inflammation. However, MES classification remains challenging due to inter-observer variability, image-quality caused label noise, and the ordinal nature of the score, which may be ignored by standard classification models. We propose CLoE, a curriculum learning framework that uses image quality for sample difficulty assignments in order to improve robustness in ordinal MES classification. Image quality, estimated via a lightweight model trained on Boston Bowel Preparation Scale (BBPS) labels, is used as a proxy for annotation confidence to order training samples from easy (clean) to hard (noisy). This quality-aware curriculum is combined with ResizeMix augmentation to enhance robustness while preserving transitions between adjacent MES grades. Experiments on the LIMUC and HyperKvasir datasets, using both CNN and Transformer backbones, show that CLoE consistently improves performance over strong supervised and self-supervised baselines. For example, ConvNeXt-Tiny achieves 82.5% accuracy and a QWK of 0.894 on LIMUC with a moderate computational footprint. These results demonstrate that difficulty-aware training guided by image quality can consistently improve ordinal disease severity assessment under label uncertainty in clinical endoscopy.Code will be released at https://github.com/zeynepozdemir/CLoE.