A Robust Vision Transformer Model for Precise Classification of Epithelial Ovarian Cancer Subtypes

The automated classification of histopathological imagery poses challenges due to diagnostic subjectivity and reliance on expert interpretation. This study presents a multitask Vision Transformer (ViT) framework designed for simultaneous subtyping of epithelial ovarian carcinoma and regression-based prediction of CA-125 biomarker concentrations from tiled Whole-Slide Images (WSIs). Leveraging GPU-accelerated computation, the model achieved 92.7 % accuracy across five subtypes-Mucinous (MC), Endometrioid (EC), High-Grade Serous (HGSC), Low-Grade Serous (LGSC), and Clear Cell Carcinoma (CC)-with a Precision of 92 %, Recall of 92 %, F1-score of 92 % and an AUC of 0.987. The framework is implemented as a Streamlit-based App development for real-time, accessible diagnostic inference. This work establishes a robust computational pathology model that enhances ovarian cancer subtype classification, supports inter-laboratory reproducibility, and contributes toward precision-driven, automated diagnostic systems.