A hybrid framework integrating deep learning and XAI for breast cancer detection with stratified cross-validation

Breast cancer remains a leading cause of mortality among women, emphasizing the critical need for early and accurate diagnosis. However, while deep learning models can achieve high accuracy in classifying mammographic images, their "black box" nature limits interpretability, making it difficult for healthcare professionals to fully trust and understand the decisions made by these models. This study introduces a hybrid framework that combines deep learning with Explainable Artificial Intelligence (XAI) to address these challenges. By incorporating XAI, the proposed framework enhances transparency, allowing key features in mammograms to be visualized, thus making the model's decisions more interpretable. Four pretrained models: VGG-16, VGG-19, ResNet50, and GoogleNet are fine-tuned and applied to classify mammographic images into three categories: benign, malignant, and normal. To mitigate data limitations, transfer learning and data augmentation are used, with performance evaluated through stratified K-fold cross-validation. Among the models, ResNet50 attained the maximum accuracy of 99.24%, along with strong metrics for AUC, precision, and F-score. Additionally, the integration of XAI techniques enables visual interpretation of the model’s decisions, helping healthcare professionals make informed and confident diagnoses. By employing XAI, the proposed framework highlights critical regions in mammograms, improving model transparency and aiding in the localization of abnormalities. This framework significantly advances diagnostics by balancing performance and interpretability, supporting radiologists in high-prevalence regions, and improving patient outcomes through actionable Artificial Intelligence-based predictions.