Visual Language Model for Preclinical Toxicologic Liver Histopathology Assessment

Preclinical drug safety assessment is a critical step in drug development that relies on time-consuming manual histopathological examination, which is prone to high inter-observer variability. Artificial intelligence methods hold promise to accelerate this assessment and enhance reproducibility. Here, we introduce VQ-TOJO, a novel self-supervised Vision Transformer model designed for general toxicologic histopathology assessment. VQ-TOJO was trained on 15 million image patches extracted from 46,734 digitized tissue sections spanning 157 preclinical studies in rats. We demonstrate VQ-TOJO's versatility in tackling various diagnostic tasks across multiple scales, including in limited labeled data scenarios. VQ-TOJO achieves state-of-the-art performance on tasks such as weakly supervised slide classification, patch-level lesion classification and automatic dose-time response assessment.