A Convolution-Attention Framework with Explainable AI for Ovarian Cancer Diagnosis using Histopathology Images

Ovarian cancer has been one of the deadliest gynecologic cancers whose survival of the patient directly depends on the initial and accurate diagnosis measure. This study presents a Convolution-Attention architecture where Explainable Artificial Intelligence (XAI) technology is applied to the automation, multi-class categorization of ovarian cancer with the help of histopathological images. The model is based on CoAtNet architecture that combines convolutional operations to extract local features to extract the global dependencies of images. System was trained using 5 histopathological categories: Clear Cell, Endometrioid, Mucinous, Serous, and Non-cancerous ovarian tissues. Emprical results reveal test reliability of 98.26 percent and a macro-average of F1-score of 0.9823 and outperforms the known CNN-based low-level models. Interpretability was also obtained using Grad-CAM visualisations which explain diagnostically meaningful tissue areas and ensure consistency between model outputs and typical pathological structures. The intersection of learning deep features and possibility of transparent visual explanations confirms the validity of the presented framework as reliable and understandable, which provides a reliable decision support system to clinical pathologists. This paper highlights the opportunities offered by having hybrid convolution-attention networks that have been augmented with explainable AI to improve the accuracy of the overall ovarian cancer diagnostics.