MammoScan: Optimizing Deep Learning Models for Breast Cancer Classification in Mammograms with Explainable AI

Abstract

Breast cancer is the most commonly diagnosed cancer worldwide and a leading cause of cancer-related deaths among women. Early detection remains the most effective way to improve patient outcomes, and mammography is widely considered the gold standard for screening. However, interpreting mammograms, particularly in women with dense breast tissue, presents significant challenges. This study aims to assist in addressing these limitations by developing a simple web application that integrates a deep learning model with explainable artificial intelligence (XAI), serving as a clinical decision support tool for clinicians and medical students. Twelve models were trained using the INBreast and CBIS-DDSM datasets, combining architectures such as ResNet50, DenseNet121, EfficientNetV2-B0, and ConvNeXt with optimizers including SGD, Adam, and AdaMax. Among these, the DenseNet121 model optimized with Adam delivered the best performance, achieving an F2 score of 67.15%, recall of 72.89%, AUC of 73.54%, precision of 51.05%, and accuracy of 63.49%. Adaptive optimizers consistently outperformed SGD, with Adam slightly outperforming AdaMax. Grad-CAM visualizations indicated that the model generally focused on clinically relevant areas, although occasional misalignment suggests the need for further refinement. The system shows potential as a clinical decision support tool in mammogram interpretation. It is intended to aid, but not replace, clinical judgment and should be used with caution.