Breast Mammogram Classification using Deep Learning

Abstract

Breast cancer remains the most common cancer among women worldwide. Early and accurate detection is vital for increasing survival rates and ensuring effective treatment planning. Despite significant progress, conventional diagnostic techniques often lack consistency and precision, leading to misdiagnosis and limited clinical reliability. To address these shortcomings, this study explores advanced deep learning approaches for multi-class classification of mammogram images into benign, malignant, and normal categories. The investigation utilized three publicly available datasets—INbreast, MIAS, and DDSM—and implemented five transfer learning models (VGG16, InceptionV3, ResNet50, and EfficientNetB0) alongside two custom Convolutional Neural Network (CNN) architectures enhanced with attention mechanisms. Among all models, the custom CNN integrated with attention achieved the highest test accuracy of 98.69%, along with F1-scores of 0.98 for both benign and malignant classifications. The transfer learning models VGG16 (93.00%) and InceptionV3 (93.90%) also yielded competitive results. Incorporating the attention mechanism significantly enhanced the network’s ability to differentiate between subtle variations in mammographic features, particularly those distinguishing benign from malignant lesions. Overall, the findings establish a reliable and high-performing framework for automated breast cancer detection, highlighting the critical role of attention-based deep learning in improving diagnostic accuracy.