Differential sparse attention framework for accurate ovarian cancer subtype classification using DenseNet121

Ovarian cancer is the most fatal gynecologic cancer, contributing to approximately 207,000 deaths per year worldwide, largely due to late-stage diagnosis and considerable histopathological diversity across subtypes. Accurate subtype classification is clinically important for prognosis and treatment planning, but remains challenging due to high inter-class morphological similarity and the scarcity of adequately labeled histopathological datasets.

This study presents a novel attention module, Deep Differential Sparse Attention (DDSA), designed for multi-class classification of ovarian cancer subtypes in histopathology images. The module selectively amplifies discriminative tumor regions while suppressing irrelevant background features, thereby improving both feature representation and model interpretability. A dataset of 3,082 images spanning five subtypes i.e., Clear Cell Carcinoma, Endometrioid Carcinoma, High-Grade Serous Carcinoma, Low-Grade Serous Carcinoma, and Mucinous Carcinoma--was obtained from Kaggle. The training set was balanced to 5,940 images (1,188 per subtype) using RandomOverSampler, with on-the-fly data augmentation applied during training. DDSA was integrated into a DenseNet-based architecture, and diagnostic performance was evaluated using standard classification metrics, including precision, recall, F1-score, confusion matrices, and ROC curves.

The proposed model achieved an overall classification accuracy of 92.53%, with a macro-average recall of 91.54% and a macro-average F1-score of 0.90. AUC values ranged from 0.92 to 0.99 across all five subtypes, indicating strong and consistent classification performance on the test set.

These results suggest that attention-guided DenseNet architectures can enhance discriminative feature extraction and improve diagnostic accuracy for ovarian cancer subtype classification. The DDSA module's ability to highlight tumor-relevant regions also supports improved model interpretability, an important consideration for clinical adoption. While these findings are promising, further validation on larger, multi-institutional datasets is needed to confirm robustness and generalizability, and to assess the framework's readiness for clinical deployment.