EXPLAINABLE BREAST CANCER DETECTION IN MAMMOGRAMS USING LIGHTWEIGHT EFFICIENTNET-B0 WITH GRAD-CAM AND LIME

Abstract

Accurate and interpretable breast cancer detection is essential for early diagnosis, yet many deep learning models remain impractical for real-world clinical deployment due to their complexity, resource demands, or lack of transparency. This study presents a comparative analysis of three CNN architectures, ResNet-50 (49.68M parameters), EfficientNet-B0 (20.36M), and EfficientNet-B7 (96.71M), to evaluate trade-offs between diagnostic performance, computational efficiency, and explainability for mammogram classification. EfficientNet-B0 achieved the highest accuracy (98.71%) and AUC (0.9996) on the MIAS dataset while requiring the fewest parameters. It also demonstrated strong generalizability on INbreast (93.32%) and DMID (94.43%) without dataset-specific fine-tuning. For interpretability, Grad-CAM and LIME visualisations were generated and qualitatively reviewed by a board-certified radiologist, who confirmed that the highlighted regionswere clinically meaningful and expressed a preference for Grad-CAM's contiguous heat-maps. These findings highlight EfficientNet-B0 as a lightweight and generalizable model, well-suited for scalable, real-world deployment in AI-assisted mammography, and underscore that robust accuracy, efficiency, and interpretability need not be mutually exclusive.