Comparative evaluation of artificial intelligence-based and pathologist scoring of murine bile duct hyperplasia.

Bile duct hyperplasia in rodents, caused by chemical cytotoxicity, bile flow interference, or aging, is traditionally assessed through time-consuming microscopic examination by pathologists. We developed a deep learning-based, artificial intelligence (AI) model for quick and accurate assessment of bile duct hyperplasia in mice. The study utilized 127 hematoxylin and eosin-stained, mouse liver whole-slide images (WSIs) digitized using Aperio scanners. A segmentation algorithm with a U-Net architecture and an EfficientNet backbone was trained on 1165 areas from 35 WSIs to identify and quantify bile ducts, measure their area percentage using a supervised method, and categorize them into grades 0 to 4 using an unsupervised method. The model was validated on 24 WSIs and tested on 68 WSIs, all independently scored by 3 veterinary pathologists using consensus criteria. The segmentation model achieved Dice scores of 91.75% in training and 85.88% in validation. Testing revealed high sensitivity (95%), specificity (97%), and Dice score (95.45%), indicating robust model accuracy. Model-predicted grades showed strong correlation (r2s = 0.86) with pathologists' consensus scores. This AI-driven approach has the potential to enhance the accuracy and efficiency of bile duct hyperplasia assessments in rodent studies, paving the way for more standardized, objective, and high-throughput histological evaluations in preclinical research.