Computational Morphometric Subtyping of High-Grade Serous Ovarian Cancer Using Routine Histopathology
TL;DR
Artificial Intelligence-derived CMBs capture clinically meaningful heterogeneity and provide a scalable New Approach Methodology (NAM) framework for immune-informed risk stratification in HGSOC.
Abstract
High-grade serous ovarian carcinoma (HGSOC) displays pronounced histological and microenvironmental heterogeneity that contributes to therapeutic resistance and poor survival. Artificial Intelligence (AI)-derived cellular morphometric biomarkers (CMBs) from routine hematoxylin and eosin (H&E) whole-slide images (WSIs) provide an unbiased means to quantify tissue heterogeneity and infer tumor microenvironment (TME) states; however, their associations with immune contexture and survival across diverse populations remain unknown. We applied a machine-learning cellular morphometric biomarker (CMB) pipeline to H&E images from 106 patients in The Cancer Genome Atlas ovarian cohort (TCGA-OV) and an independent cohort of 22 patients with HGSOC collected at Loma Linda University (LLU-OV). Differential CMBs were identified, corroborated across cohorts, and evaluated for associations with overall survival (OS). Immune deconvolution and checkpoint gene expression analyses were performed using TCGA data, while CD3, CD8, and PDCD1 immunohistochemistry (IHC) assessed immune infiltration in LLU samples. Three reproducible CMBs (CMB73, CMB80, CMB215) demonstrated consistent patterns across cohorts. Higher abundances of CMB73 and CMB80 were associated with worse OS and reduced immune infiltration, whereas CMB215 correlated with improved OS and immune-enriched TMEs, including increased PDCD1, PDCD1LG2, and CD8A expression. IHC findings showed significant association of OS with T cells and age. Overall, AI-derived CMBs capture clinically meaningful heterogeneity and provide a scalable New Approach Methodology (NAM) framework for immune-informed risk stratification in HGSOC.