Explainable Lightweight Deep Learning with Grad-CAM for Breast Cancer USG Classification

Abstract

Breast cancer is still one of the most common cancers worldwide, with about 2.26 million new cases in 2020. Ultrasound (USG) images are an important screening method because they are non-invasive and affordable. However, USG images often have speckle noise and low contrast, which can make diagnosis less consistent. While deep learning has significantly improved classification accuracy, many complex models require substantial computational resources, limiting their application on portable or edge devices. Furthermore, the opaque nature of these models often hinders clinical trust and interpretability. In this study, we propose a precise and interpretable classification system for breast ultrasound images using a combined dataset of 1,030 samples from the BUSI and Mendeley repositories. We evaluated six distinct deep learning architectures: DenseNet121, EfficientNetB3, EfficientNetB4, MobileNetV2, MobileNetV3, and VGG16. To address the black box nature of deep learning, we integrated the Grad-CAM technique, which generates visual saliency maps to highlight relevant pathological features. Our experiments showed that MobileNetV3 had the best overall performance, with 92.41% accuracy and a 91.19% F1-score. EfficientNetB3 had the highest sensitivity, with a recall of 96.98%. Grad-CAM visualizations showed that the models focused on the important areas of the lesions. These results suggest that using lightweight models with Explainable AI (XAI) can offer an effective, reliable, and efficient diagnostic tool for real-world clinical use, especially where resources are limited.