An Explainable Transfer Learning Framework for Breast Cancer Classification Using Deep CNN Models

Abstract

When it comes to cancers affecting women, breast cancer is still at the top of the list and a huge problem for public health. The key to better patient outcomes and lower mortality rates is early and precise diagnosis. In order to overcome these obstacles, this research presents a hybrid framework that uses deep learning (DL) in conjunction with XAI. The proposed framework's use of XAI improves transparency, which in turn makes the model's judgments easier to understand by visualizing important aspects in mammograms. Classification of mammographic pictures into benign, malignant, and normal is achieved through the application of four pretrained models: GoogleNet, ResNet50, VGG-19, and VGG-16. Stratified K-fold cross-validation is utilized to evaluate performance as a means to address data limitations through the use of data augmentation and transfer learning. Of all the models, ResNet50 achieved the highest possible accuracy of 99.34% and performed admirably in terms of area under the curve (AUC), precision, and F-score. Visual interpretation of the model's decisions is also made possible by XAI approaches, which aids healthcare professionals in making confident and informed diagnoses. Improving model transparency and aiding in the identification of anomalies, the proposed framework utilizes XAI to highlight key locations in mammograms. Supporting radiologists in high-prevalence locations and increasing patient outcomes with actionable AI-based anticipations, this approach considerably advances diagnostics by balancing performance and interpretability.