Interpretable Deep Learning for Ovarian Cancer Subtype Classification Using Histopathology Images

Abstract Ovarian cancer is a histologically heterogeneous malignancy in which accurate subtype classification is essential for prognosis and treatment selection. Manual interpretation of hematoxylin and eosin (H&E)–stained histopathology slides remains time-consuming and subject to inter-observer variability, particularly for morphologically overlapping subtypes. In this study, an interpretable deep learning framework was developed for automated ovarian cancer subtype classification using a fine-tuned ResNet50 architecture. A publicly available histopathology dataset comprising five major ovarian carcinoma subtypes was employed. Model training incorporated optimized strategies including data augmentation, selective layer unfreezing, label smoothing, and test-time augmentation. Classification performance was benchmarked against established convolutional neural network architectures. Visual interpretability was assessed using Gradient-weighted Class Activation Mapping (Grad-CAM) to examine model attention patterns. The proposed framework achieved high classification performance, with strong accuracy and macro-averaged F1-scores across subtypes. Grad-CAM analysis demonstrated that model attention was consistently localized to histologically relevant tissue regions in correctly classified samples, while misclassifications were associated with morphologically ambiguous structures. These findings suggest that classification errors primarily reflect intrinsic histopathologic overlap rather than spurious model behavior. Overall, this work establishes a reproducible and interpretable deep learning benchmark for ovarian cancer subtype classification and provides clinically meaningful insight into model decision processes. The proposed framework is well-suited for use as a second-reader decision support tool in digital pathology workflows.