Dual-branch CrossViT for ovarian cancer diagnosis: Integrating and explainable AI for real-time clinical applications

Early and accurate detection of ovarian cancer significantly improves patient outcomes by allowing for timely treatment. This study introduces a deep learning (DL) framework using a dual-branch Cross Vision Transformer (CrossViT) for classifying ovarian cancer subtypes through high-resolution histopathological images. Unlike traditional convolutional neural networks (CNNs), which struggle with capturing global dependencies, CrossViT utilizes multi-scale self-attention to extract detailed textural patterns and broader contextual information. This design addresses class imbalance and enhances feature learning, leading to improved diagnostic accuracy. A dataset of 100,000 histopathological images representing five ovarian cancer subtypes was compiled from Kaggle. The images underwent preprocessing, including noise reduction, data augmentation to balance class sizes, and pixel normalization for uniformity. The model also uses Gradient-weighted Class Activation Mapping (Grad-CAM) to highlight important image regions for classification, ensuring transparency and clinical reliability. Results show that the CrossViT model outperforms existing CNN models, achieving a classification accuracy of 99.24% and superior scores in F1, specificity, Matthews Correlation Coefficient (MCC), and Precision-Recall AUC (PR AUC). Additionally, a real-time web application has been developed for clinicians to quickly classify subtypes from histological samples. Future work will focus on improving computational efficiency and using more diverse datasets to enhance generalizability and clinical use.