LIGHTWEIGHT RESNET-18 BASED DEEP LEARNING FRAMEWORK FOR BREAST CANCER DETECTION USING MAMMOGRAPHY

Breast cancer is one of the major types of cancer affecting women globally. Mammography is a standard modality for early detection due to its accessibility, economic viability, and proven diagnostic reliability; however, its interpretation remains a difficult task due to factors such as heterogeneous breast density, overlapping fibro-glandular structures, subtle morphological abnormalities, and significant inter-reader variability among radiologists with differing levels of expertise. These challenges highlight the need for automated, objective, and reproducible methods to enhance breast cancer detection and reduce diagnostic inconsistencies. In this study, we introduce a lightweight, computationally efficient deep learning framework based on a modified ResNet-18 architecture designed specifically for mammography classification (cancer vs. normal). Through a combination of global average pooling, residual feature extraction, and supervised training using cross-entropy loss, the framework aims to provide robust prediction performance while also generating interpretable Grad-CAM heatmaps that highlight regions most influential to the model’s decision-making. The proposed approach can be deployed in clinical workflows, offers near real-time inference speed, and is suitable for evaluation on standard mammography datasets using metrics including sensitivity, specificity, accuracy, and receiver operating characteristic (ROC) analysis. The proposed framework provides a clinically relevant, transparent, and computationally efficient foundation for improving breast cancer screening outcomes and supporting radiologists in diagnostic decision-making.