An AI system trained to read mouse liver slides can flag toxic tissue damage — including pathology types it has never seen before — with a false negative rate of just 0.16 percent.
Researchers built the framework around a fine-tuned Vision Transformer (DINOv2), adapted using Low-Rank Adaptation (LoRA) and trained on a newly assembled dataset of pixel-annotated rodent liver slides. Once trained, the system classifies tissue at the pixel level and uses Mahalanobis distance — a statistical measure of how far a data point sits from a known distribution — to flag samples that fall outside its training categories. That out-of-distribution capability matters: the model was tested on apoptosis and staining artifacts, two tissue states it had never been trained on, and still correctly identified 89.38 percent of those novel findings.
Drug-induced toxicity is one of the most common reasons compounds fail in preclinical and early clinical trials, and the current standard — expert pathologist review of histopathology slides — does not scale to the volume of candidates that modern drug pipelines generate. A system that can screen whole-slide images automatically, and that errs decisively on the side of not missing lesions, could shrink both the time and cost of that bottleneck.
The results are a proof of concept on mouse liver tissue, not a cleared clinical tool, and rodent slides are a narrower domain than the full range of tissues and species a drug development program touches — but the architecture's ability to generalize to unseen pathology types is the detail worth watching.