Unsupervised Single-Class Anomaly Detection in Rat Liver Using a Model of Normal Histomorphology.

An unsupervised, single-class anomaly detection approach based on a vision transformer architecture was developed to aid histopathology evaluation of rat liver from nonclinical toxicology studies. The approach was designed to detect any histopathology finding without requiring specific histopathology examples for training. The model was developed by training a vision transformer with approximately 135,000 whole slide images (WSIs) and then extracting high-dimensional feature representations from 750 unremarkable vehicle control WSIs, creating a rat liver histomorphology reference dataset representing typical biological variability. Each test WSI was compared against the reference dataset at the tile level, and the resulting anomaly detections were aggregated to derive an overall image-level anomaly score. Anomaly scores were higher in affected groups treated with established toxicants than in control groups, and based on comparing anomaly scores at the individual animal level with pathologist diagnoses, model sensitivity was 95.9% and specificity was 70.8%. Utility as a pathologist aiding tool was further studied by comparing tile-level anomaly annotations with diagnosis. For one treated group, histomorphology feature extraction was used to explore discordance between anomaly detection and diagnosis. Specificity was impacted by pre-analytical variables as well as detection of anomalies not considered noteworthy by pathologists.