Identifying Ovarian Cancer Subtypes in a Multiple-Instance Learning Approach

Ovarian cancer is one of the most lethal gynecological malignancies, where accurate subtype identification through the histopathological analysis of whole slide images (WSIs) is critical for diagnosis and treatment. In this paper, we propose a Multiple Instance Learning (MIL) framework that integrates Transformer architectures to improve the precision of WSI-based identification for ovarian cancer subtypes. The proposed framework processes WSIs by dividing them into small tissue patches and employs a Transformer model to analyze spatial relationships and contextual patterns across these patches. We further enhance this TransMIL model’s capability by incorporating a MultiScale Global Attention (MOGA) module, which improves the integration of both broad tissue structures and localized pathological details. Our experimental evaluations demonstrate that the TransMIL-MOGA model surpasses traditional MIL approaches, such as mean-pooling, max-pooling, TransMIL and the AttentionBased MIL (ABMIL) model, in achieving consistently higher accuracy, recall, and F1-scores. The MOGA module proves particularly effective in synthesizing multi-scale information, enabling the model to distinguish subtle histological differences between ovarian cancer subtypes. This TransMIL-MOGA model may serve as an effective decision-support tool for pathologists on ovarian cancer diagnosis on handling class-imbalanced medical datasets and improving diagnostic consistency.