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

Area of Science:

• Digital pathology
• Artificial intelligence in medicine
• Gastroenterology

Background:

• Histopathological assessment is crucial for inflammatory bowel disease (IBD) diagnosis and management.
• Standardized scoring systems like the Nancy Histological Index (NHI) have limitations in observer reproducibility.
• Digital pathology offers potential for automated and consistent histological analysis.

Purpose of the Study:

• To develop and evaluate REMIL-IBD, a weakly supervised framework for automated IBD inflammation grading using whole-slide images.
• To leverage pretrained histology foundation models for accurate and interpretable histopathological assessment.
• To reduce the need for detailed region-level annotations in clinical settings.

Main Methods:

• REMIL-IBD utilizes a weakly supervised, multiple instance learning approach on hematoxylin and eosin-stained biopsies.
• The framework incorporates attention-based learning and pretrained histology foundation models (UNI, Virchow2, Cerberus).
• Evaluation involved both five-class NHI grading and three-class disease activity groupings.

Main Results:

• REMIL-IBD achieved approximately 82-84% accuracy in three-class classification tasks.
• The framework demonstrated strong discrimination between histological remission and severe active disease.
• Region filtering improved classification for higher inflammation grades, enhancing quantitative tissue feature extraction.

Conclusions:

• Pretrained histology foundation models enable reliable and interpretable grading of IBD inflammation with limited annotation.
• The REMIL-IBD framework offers a scalable and clinically relevant solution for automated histopathological assessment in digital pathology.
• This approach supports practical deployment in real-world clinical settings for IBD management.