Detection of macrovesicular steatosis in hematoxylin and eosin-stained histological images of human livers: A feature-based method

Macrovesicular steatosis (MaS) affects liver transplant outcomes. Traditional visual biopsy assessment is subjective and shows inter-observer variability (weighted κ = 0.595 in our cohort), which complicates allocation and prognostication. We developed a semi-automated image analysis method using HALCON Progress (evaluation license, 2019–2023) to quantify MaS in H&E whole-slide images (WSIs). The pipeline processes native SVS files at full resolution (average runtime 1.26 ± 0.53 min/WSI) without tiling, downsampling, or format conversion. Four shape features (area, roundness, circularity, and compactness) guide classification. Results were compared with three specialized pathologists using Pearson and Spearman correlations. Across 129 WSIs (≈52,000 lipid droplets), artificial intelligence–pathologist correlations were statistically significant (Pearson r = 0.526–0.642; Spearman ρ = 0.498–0.615; all p < 0.001; n = 48–95 per pathologist). Correlation with the mean pathologist assessment reached R 2 = 0.64 ( r = 0.80, ρ = 0.782), within the inter-pathologist range (R 2 = 0.34–0.62; weighted κ = 0.595). Using four shape features, the pipeline separates vacuoles from vessels and processing artifacts during interactive review. Processing native SVS at full resolution avoids extra compute while preserving precision. The method provides a rapid, objective readout of MaS and performs on par with inter-pathologist agreement (R 2 = 0.64 vs. weighted κ = 0.595). It is best used as decision-support that supplies percentages and overlays for expert review. Main limitations are underestimation in severe, confluent steatosis (>30%), limited coverage of microvesicular fat, and single-center validation. Multi-center evidence across scanners and patient groups is still needed before routine use.