OvarianNet-Ca: A Hybrid Cross-Attention Ensemble Model Approach Using MixTransformer and EfficientNet
Deepan Adak, Srushti Sonawane, Gaurav Verma · 2025
Ovarian cancer classification poses significant challenges in pathological diagnosis due to its complexity and the critical need for accurate identification for targeted treatment. This research introduces OvarianNet-Ca, a novel hybrid cross attention based ensemble architecture that uses EfficientNetV2 and MixTransformer models through a cross-attention mechanism to facilitate the accurate classification of ovarian cancer subtypes. In order to efficiently manage multi-institutional variations in histopathological images, the work implements a custom normalization block and hierarchical feature fusion strategy. The UBC-OCEAN dataset, which includes whole slide images and tissue microarrays from more than 20 institutions on four continents, was used to evaluate the model. OvarianNet-Ca achieved state-of-the-art performance with 95.221% ROC-AUC, 97.621% accuracy, and 95.103% F1-score across different imaging conditions. The effectiveness of each architectural component is validated by comprehensive ablation studies, which show the importance of the cross-attention mechanism and custom normalization block in managing intricate histopathological patterns. The model’s potential to enhance diagnosis accuracy and accessibility in clinical settings is promising, especially in areas deficient in expert gynecologic pathologists. This study signifies a significant progress in computational pathology, offering a reliable approach for the automated classification of ovarian cancer subtypes that sustains high efficacy across various healthcare environments.