Deep Learning Models Capture Histological Disease Activity in Crohn's Disease and Ulcerative Colitis with High...

Area of Science:

• Gastroenterology
• Computational Pathology
• Artificial Intelligence

Background:

• Histological disease activity is crucial for clinical outcomes and drug development in inflammatory bowel disease (IBD).
• Automating histological assessments in IBD is essential for efficient and objective evaluation.

Purpose of the Study:

• To develop and evaluate deep learning models for automating histological assessments in IBD.
• To compare the performance of different AI methods in predicting histological disease activity scores.

Main Methods:

• Histology images from Crohn's disease (CD) and ulcerative colitis (UC) clinical trials were used to train AI models.
• Models predicted the Global Histology Activity Score (GHAS) for CD and Geboes score for UC.
• AI model predictions were compared against expert central readings and independent pathologists.

Main Results:

• A self-attention-based multiple instance learning model (SA-AbMILP) showed superior performance.
• AI-predicted GHAS and Geboes subgrades demonstrated moderate to substantial agreement with central readings (65-89% accuracy).
• The model accurately distinguished histological features in CD and UC across different anatomical locations (76-94% accuracy).

Conclusions:

• Deep learning models effectively distinguish microscopic disease activity in IBD based on established scoring systems.
• AI-driven histological assessment shows comparable accuracy to human pathologists, supporting its use in clinical trials.