Interpretable predictions of toxicological liver abnormalities through multiple instance learning algorithms applied to digital whole slide images

Study Background

The examination of toxic effects on the liver and other organs in mammalian preclinical models is essential to understand the safety
profile of experimental therapeutics.

The identification of potential toxicities is made through the assessment of histologic remarkability – the presence of abnormalities such as necrosis, inflammation, or pigment accumulation – on histologic slides. However, the accurate detection of these events can be hindered by inter-pathologist variability and the potential subtlety of meaningful changes in tissue architecture.

To explore the utility of incorporating artificial intelligence into toxicologic pathology evaluations, we developed an algorithm to predict the presence of abnormalities in hematoxylin and eosin (H&E)-stained specimens.

Conference

Society for Toxicologic Pathology 2025

Partner

Merck