AI-driven multi-scale histopathologic classification of epithelial ovarian cancer to support precision oncology.

Abstract

5608 Background: Epithelial ovarian cancer (EOC) comprises five histologically distinct subtypes associated with divergent molecular profiles, therapeutic responsiveness, and prognostic outcomes. Accurate histologic classification is essential for precision oncology but remains limited by morphologic overlap and interobserver variability. Deep learning architectures capable of multi-scale spatial feature extraction may enable objective and reproducible subtype discrimination. We evaluated the performance of a multi-scale factorized convolutional neural network (InceptionV3) for automated EOC subtype classification. Methods: Digitized hematoxylin–eosin–stained histopathology images were obtained from the University of British Columbia Ovarian Cancer Subtype Classification and Outlier Detection (UBC-OCEAN) dataset, including high-grade serous, low-grade serous, endometrioid, clear cell, and mucinous carcinomas. An ImageNet-pretrained InceptionV3 architecture was fine-tuned using a five-class classification head. The model leverages parallel convolutional branches with factorized asymmetric kernels to simultaneously capture fine cellular features and global tissue architecture. Images were normalized and augmented. Training was performed using 299×299 inputs, Adam optimizer (learning rate 1×10⁻⁴), cross-entropy loss, and six epochs. Performance was assessed using accuracy, precision, recall, F1-score, and macro-averaged AUROC. External clinical validation was conducted by 57 board-certified gynecologic pathologists and oncologists across 14 geographic regions. Results: The model achieved an overall accuracy of 95.0% (491/517 test images). Subtype-specific F1-scores were 0.96 for high-grade serous carcinoma, 0.94 for endometrioid, 0.93 for clear cell, 0.92 for mucinous, and 0.88 for low-grade serous carcinoma. Macro-averaged AUROC was 0.957 (95% CI, 0.948–0.966). Computational efficiency was preserved with 23.9 million parameters and 11.6 billion FLOPs. Independent clinician validation demonstrated strong concordance with AI predictions, with mean inter-rater agreement of 95.3% (κ=0.936; 95% CI, 0.923–0.949), consistent across institutional practice settings. Conclusions: Multi-scale factorized convolutional modeling enables near-expert performance for automated EOC subtype classification by integrating cellular- and tissue-level morphologic features. High inter-rater concordance with an international specialist cohort supports translational reliability. This framework represents a scalable decision-support tool with potential to improve diagnostic standardization and precision therapeutic stratification. Prospective multi-center validation is warranted.