Machine learning-enabled quantification of hepatocellular necrosis in the liver after lethal Marburg and Ebola virus exposures.

The liver is a key target of pathogenic orthomarburgviruses and orthoebolaviruses, with injury common in severe filovirid disease, yet high-resolution histopathologic quantification is lacking. We addressed this by deploying deep learning (DL) models for pixel-level quantitative analysis of digitized liver pathology slides in lethal rhesus monkey (RM) models of Marburg virus (MARV) and two Ebola virus (EBOV) variants, Makona and Kikwit. Our DL model segmented arteries, veins, bile ducts, and hepatic necrosis, achieving interobserver variability in necrosis segmentation comparable to three pathologists. DL-quantified liver necrosis correlated with exposure virus (f=6.61, p=.006) and was highest in MARV-exposed RMs (11.0%). While filovirid-exposed RMs showed elevations in aspartate aminotransferase (AST), alanine aminotransferase (ALT), gamma-glutamyltransferase (GGT), and alkaline phosphatase (ALP), only ALT levels correlated with necrosis severity. Necrosis localization differed: portal tract-proximate after MARV exposure and central vein-proximate after EBOV exposure. This proof-of-concept work enables future large-scale retrospective "meta-pathologic" analyses.