A Multi-Task Vision Transformer Framework for Automated Classification of Epithelial Ovarian Cancer Subtypes

Abstract

Classification of medical images is a difficult and time consuming job that requires a specialist's eye and is prone to errors, making an automated system vital. This study presents a robust multi-task Vision Transformer (ViT) model for the joint classification of epithelial ovarian cancer subtypes and prediction of CA-125 biomarker levels from histopathology images. The model achieved a state-of-the-art overall accuracy of 95.7% in classifying five major subtypes: Mucinous Carcinoma (MC), Endometrioid Carcinoma (EC), High-Grade Serous Carcinoma (HGSC), Low-Grade Serous Carcinoma (LGSC) and Clear Cell Carcinoma (CC). Performance metrics demonstrated exceptional robustness, with precision 96.2%, recall 95.1% and F1-scores 95.4%. A near-perfect overall AUC of 0.992 in ROC analysis confirmed superior diagnostic capability across all classes. The confusion matrix revealed minimal, clinically understandable misclassifications. The training trajectory showed rapid convergence and optimal generalization without overfitting. These results signify a major advancement in computational pathology, providing a highly accurate, automated tool for ovarian cancer subtyping that can augment pathological diagnosis, improve standardization and potentially support personalized treatment strategies.