5520 Background: Despite serving as the first-line systemic treatment for high-grade serous ovarian cancer, roughly 20% of patients fail to respond to platinum-based chemotherapy. However, there is a paucity of biomarkers which correlate to platinum resistance. Effective biomarkers would allow for individualized risk profiling and early identification of patients who may benefit from treatment escalation. Quantitative pathologic biomarkers have shown prognostic utility in several cancers and offer a potential low-cost approach to address challenges in assessing platinum resistance. We examined the utility of machine learning (ML)-derived pathologic biomarkers in identifying platinum-resistant ovarian cancer. Methods: Our multi-institutional dataset included 158 whole-slide images (WSIs) obtained from resection of primary ovarian cancer, with 86 (54%) in the training cohort and 72 (46%) in the validation cohort. Our dataset included 112 (71%) platinum-sensitive and 46 (29%) platinum-resistant WSIs. The CLAM deep learning framework and CONCH visual-language foundation model were utilized to segment tissue, generate 256 x 256 image patches, and extract 512-dimensional features from each patch. Non-linear dimensionality reduction, employing Uniform Manifold Approximation and Projection (UMAP), was applied to cluster training samples, and validation samples were projected onto the learned low-dimensional space. Samples were stratified into predicted-sensitive and predicted-resistant cohorts based on similarity to training-derived clusters, and clinical characteristics were compared using a two-tailed t-test and chi-square tests. Results: UMAP analysis identified distinct phenotypes associated with platinum sensitivity and resistance. At the image level, UMAP-defined phenotypes exhibited high predictive ability. Of the predicted-sensitive images in the validation cohort, 79% showed platinum sensitivity. UMAP-defined phenotypes demonstrated moderate patch-level accuracy in identifying platinum-sensitive samples (training accuracy 69%, validation accuracy 65%). Additionally, clinical correlates to pathology-based predictions were investigated. Predicted-resistant patients were significantly older (median age, 69 years IQR 59.5–76.26 vs. 62 years IQR 55–68, p = .021) and more frequently higher-staged (IIIC/IV 89% vs. 77%, p = .087) compared to predicted-sensitive patients. Conclusions: ML-derived pathologic biomarkers can aid in the identification of platinum-resistant ovarian cancer. Our future work will evaluate the relationship between quantitative biomarkers and known clinical predictors of platinum resistance, such as CA-125 and the extent of residual disease at cytoreduction. Integrating multimodal biomarkers in clinical decision-making practices will allow for improved risk stratification and individualized treatment for patients.
Sritharan et al. (Wed,) studied this question.