Quantitative fibrosis features from digitized H&E-stained liver biopsies reveal granular insights into fibrosis and disease progression in metabolic dysfunction-associated steatohepatitis: a retrospective analysis of STELLAR-3 and STELLAR-4

Background Histologic staging of metabolic dysfunction-associated steatohepatitis (MASH) requires semiquantitative assessment of hepatocellular ballooning, steatosis, lobular inflammation, and fibrosis and is fraught with pathologist variability. We hypothesize that quantitative histologic analysis will better reflect the continuous distribution of histologic features, and thus the disease biology. Methods Using LiverExplore, a suite of machine learning models that detect and classify liver tissue regions, lobular zones, cell types, and fibrosis subtypes from hematoxylin and eosin-stained whole slide images, we characterized the MASH disease microenvironment in participants of the STELLAR-3 and STELLAR-4 trials ( NCT04052516 ). The correlation of model-derived human interpretable features (HIFs) with pathologist-provided MASH CRN grades and fibrosis stages, AI-generated continuous CRN grades and stages, non-invasive biomarkers, transcriptomics, and patient outcomes was assessed. Results Model-quantified steatosis, lobular inflammation, hepatocellular ballooning, and fibrosis features were significantly correlated with pathologist CRN grades/stages, while model-derived tissue and cell features revealed quantitative changes in the disease microenvironment as MASH progressed. Pathological and advanced fibrosis HIFs were correlated with non-invasive metrics of fibrosis and a gene signature associated with hepatic stellate cells. HIFs associated with nodular or advanced fibrosis and inflammation were associated with an increased risk of liver-related events in patients from STELLAR-3 and STELLAR-4. Conclusions AI-powered quantitative characterization of the liver microenvironment delivers context relevant to MASH progression beyond the resolution afforded by categorical CRN scoring, highlighting the promise of this approach for broad applications in drug development.