REMIL-IBD: Region-filtered multiple instance learning for interpretable slide-level grading of inflammatory bowel disease

Histopathological assessment of inflammatory bowel disease (IBD), including ulcerative colitis and Crohn's disease, is essential for diagnosis, disease monitoring, and treatment planning. Standardized scoring systems such as the Nancy Histological Index (NHI) provide structured measures of disease activity; however, reproducibility can vary across observers, particularly for intermediate grades of inflammation. As digital pathology continues to expand in clinical practice, it opens new possibilities for automated, consistent, and interpretable assessment of histological disease activity using routine clinical data. In this study, we propose REMIL-IBD (Region-Filtered Embedding-based Multiple Instance Learning), a weakly supervised framework for automated grading of inflammation directly from whole-slide images of hematoxylin and eosin-stained biopsies. The proposed approach leverages pretrained histology foundation models within an attention-based learning framework to identify diagnostically relevant tissue regions and generate interpretable slide-level predictions from routine clinical data. By relying only on slide-level labels, the method reduces the need for detailed region-level annotations and supports practical deployment in real-world clinical settings. We evaluated three REMIL-IBD variants using publicly available foundation models (UNI, Virchow2, and Cerberus) across conventional five-class NHI grading and clinically motivated three-class disease activity groupings. The proposed framework achieved accuracies of approximately 82-84% in three-class classification tasks, with particularly strong discrimination between histological remission and severe active disease. Region filtering enables the extraction of quantitative tissue features and gives improved classification performance for higher-severity inflammation grades, while introducing additional preprocessing time per slide. The results demonstrate that pretrained histology foundation models can enable reliable, interpretable grading of IBD inflammation using lightweight downstream models and limited annotation. The REMIL-IBD framework provides a scalable and clinically relevant approach for automated histopathological assessment of IBD in digital pathology workflows.