Abstract B039: Application of unsupervised clustering to whole slide images of high-grade serous ovarian cancer predicts prognosis

Abstract Introduction: High-grade serous ovarian cancer (HGSOC) has a median survival < 5 years. Yet 10% of patients live > 10 years despite similarly aggressive initial presentation and receipt of the same cytoreductive and platinum-containing 1st line treatment. Understanding the extremes of survival may reveal insight into HGSOC biology, however there are no reliable means of predicting prognosis. Application of artificial intelligence (AI) to hematoxylin and eosin-stained whole slide images (H&E WSIs) have attained modest results in predicting clinical outcomes across cancer types, however these promising methods are limited in their interpretability and reproducibility. We developed an AI model to predict prognosis of patients with HGSOC from H&E WSIs and analyzed the metadata generated to drive understanding of both model function and the contribution of histologic features to survival. Methods: Survival categories were defined: short-term survival (STS) as < 2 years, long-term survival (LTS) as > 7 years, and medium-term survival (MTS) as between 2-7 years. From a cohort of 150 patients (272 H&E WSIs – 81 STS, 128 MTS, 63 LTS) with HGSOC who underwent primary or interval debulking surgery and adjuvant platinum-containing chemotherapy, we extracted features from the H&E WSIs at 10x magnification using the CONCH foundation model. We applied PhenoGraph Leiden community detection, a method of unsupervised clustering, to characterize the cohort at the level of these 224 x 224-pixel tiles. The resulting clusters were evaluated by trained pathologists and independently correlated to survival via Cox proportional hazards model. We then compared simple machine learning methods to predict survival category based on composition of each H&E WSIs with respect to cluster. Results: We identified 45 morphologically distinct clusters of H&E features that are interpretable to pathologists, finding several clusters to be associated with the ends of survival. When visualized on a Uniform Manifold Approximation Projection (UMAP), the clusters exhibit an identifiable spectrum based on tile composition. The best performing predictive model achieved an average area under the receiver operating curve of 0.757 with sensitivity of 0.567 and specificity of 0.795 for this 3 category survival classification. Conclusions: We present a method that refines abstract, foundation model-based features from H&E WSI tiles into interpretable clusters of morphologically similar tiles. These clusters are independently associated with survival and can effectively be utilized to predict prognosis. Grouping tiles into visually coherent clusters removes a layer of opacity from the “black box” aspects of the AI model. Further work is needed to optimize the predictive model and validate the generalizability of this method. Citation Format: Kelsey C. Goon, Leonhard Donle, Ernst Lengyel. Application of unsupervised clustering to whole slide images of high-grade serous ovarian cancer predicts prognosis [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 A039.