Explainable MobileNetV2 Model for Different BI-RADS Breast Cancer Diagnosis Using Mammogram Scans
Israa Abdelsabour, Ahmed Elgarayhi, Mohammed Sallah, Mohammed Mahfouz Elmogy · 2025
Breast cancer (BC) is the most powerful cause of mortality among women worldwide. Mammography is the primary screening tool for early recognition; however, the accurate discovery and classification of malignant lesions remain a challenge for systems. Deep learning (DL) models have shown the potential to improve computer-aided diagnosis (CAD) systems. This study compares the effect of different DL models, such as MobileNetV2, Inception-ResNetV2, and DenseNet121, to classify mammogram images using the King Abdulaziz University Breast Cancer Data Set (KAU-BC). The dataset comprises six BI-RADS categories; we focus on four key classes: normal (BI-RADS 1), probably benign (BI-RADS 3), suspicious malignant (BI-RADS 4), and malignant (BI-RADS 5). MobileNetV2 is a particularly acceptable pre-trained model for resource-limited settings, such as hospitals and mobile applications, due to its lightweight architecture. To enhance model interpretability, gradient-weighted class activation mapping (Grad-CAM) is utilised, facilitating visual explanations by highlighting essential regions that affect the classification decision. The proposed approach gives high performance with MobileNetV2, with a general cross-validation accuracy of$\mathbf{9 6. 8 \%}$, a sensitivity of$\mathbf{9 6. 8 \%}$, a specificity of$\mathbf{9 8. 9 \%}$, and an F1 score of$\mathbf{9 6 \%}$for the diagnosis of BC.