Pediatric Inflammatory Bowel Disease Tissue Classification From Pathology Slide Images: Detecting Phenotypes Using...

Area of Science:

• Digital pathology
• Computational pathology
• Artificial intelligence in medicine

Background:

• Histopathology images from endoscopic biopsies are crucial for diagnosing inflammatory bowel disease (IBD).
• Automated classification of histologic phenotypes could guide Crohn's disease and ulcerative colitis diagnosis and treatment.
• Artificial intelligence (AI) may improve pediatric IBD outcomes by assisting pathologists.

Purpose of the Study:

• To assess AI's capability in classifying histopathology images for pediatric IBD.
• To evaluate AI's accuracy in detecting abnormal tissue sections and specific disease features.

Main Methods:

• Development of three 2D convolutional neural networks (CNNs) utilizing multiple instance learning.
• Training models to classify histopathology slides as normal vs. abnormal.
• Training models to detect active inflammation and chronic changes/architectural distortion.

Main Results:

• Abnormal vs. normal classification achieved 0.84 accuracy and 0.91 AUC-ROC.
• Active inflammation prediction reached 0.85 accuracy and 0.92 AUC-ROC.
• Chronic changes prediction showed 0.86 accuracy and 0.93 AUC-ROC.

Conclusions:

• A significant AI-interpretable signal exists in endoscopic whole slide imaging.
• Weakly supervised methods like multiple instance learning are effective.
• AI holds potential to enhance diagnostic accuracy in pediatric IBD.