Abstract B040: Applying histopathology foundation model-enhanced neural networks to ovarian tumor whole slide images and tissue microarray for survival prediction and tumor microenvironment subtyping

Background: Artificial intelligence-based histopathology image analysis using whole slide images (WSIs) has rapidly expanded in recent years with wide implications and opportunities for predicting clinical outcomes. Here we applied foundation model-based features to hematoxylin and eosin (H&E)-stained ovarian tumor WSIs and tumor microarrays (TMAs) and trained neural networks (NN) to predict survival and classify RNA-based tumor microenvironment (TME) subtype. Methods: We developed survival models for epithelial ovarian cancers using NNs enhanced with foundation model features from H&E-stained TMA and WSI across four cohorts (n=1,453). Tile-level features were extracted at 20x with foundation models: CHIEF, GIGAPATH, VIRCHOW, and input into a Weibull-based deep survival model (PyTorch). Models were trained separately on the TCGA (WSI) and the New England Case Control Study (NECC, TMA) using k-fold cross-validation with age-stratified sampling. Performance was assessed using held-out validation sets from TCGA and NECC, and externally on two independent datasets. TME subtypes were defined by bulk RNA-seq data based on cancer-associated fibroblasts (CAF), T cells, and endothelial cells. Unsupervised k-means clustering (k=4) with Euclidean distance over 1000 bootstrapped iterations was performed, and each cluster was annotated by the median expression scores of the selected marker sets and as one of the following subtypes:1) Immune-Enriched, 2) Immune-Enriched + Fibrotic, 3) Fibrotic, or 4) Depleted. We applied the classification NN for slide-level (multilayer perceptron) and core-level classification (Attention-Based Deep Multiple Instance Learning; ABMIL) to train models on TMA-derived H&E tiles and classify samples into the four transcriptomic-defined subtypes. Results: Survival models trained on TMAs and WSIs achieved comparable performance across validation cohorts with strong prediction performance (c-index=0.76-0.96). In both NECC and TCGA, Immune-Enriched + Fibrotic and Depleted clusters were most distinct, marked by high and low expression of CAF, T cell, and endothelial scores, respectively. In contrast, Immune-Enriched and Fibrotic clusters were more difficult to differentiate, particularly based on CAF and endothelial scores, though T cell scores were higher in Immune-Enriched relative to Fibrotic clusters. The TMA-trained NN models were unable to reliably distinguish between the four RNA-defined TME subtypes. Conclusion: Despite TMA’s reduced spatial context, survival prediction was comparable to WSIs with consistent performance in validation datasets, suggesting that TMA-derived features paired with foundation models can capture biologically important signals in epithelial ovarian tumors. RNA-seq-defined TME subtype classification using TMA-derived H&E features was limited, likely due to sample size, spatial resolution, and the heterogeneity of the ovarian cancer TME. Future direction will leverage attention-based models to generate spatial maps highlighting tissue regions predictive of molecular features and clinical outcomes. Citation Format: Tina Yi Jin Hsieh, Shih-Yen Lin, Allison F. Vitonis, Kathryn L. Terry, Kun-Hsing Yu, Naoko Sasamoto. Applying histopathology foundation model-enhanced neural networks to ovarian tumor whole slide images and tissue microarray for survival prediction and tumor microenvironment subtyping [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Advances in Ovarian Cancer Research; 2025 Sep 19-21; Denver, CO. Philadelphia (PA): AACR; Cancer Res 2025;85(18_Suppl):Abstract nr B040.