A Comparative Study of Deep Learning Models and XAI for Breast Ultrasound Image Classification
Muhammad Rakha, Adiwijaya Adiwijaya, Untari Novia Wisesty, Putu Harry Gunawan · 2025
Breast cancer continues to be a major cause of death among women globally. Early diagnosis greatly enhances the chances of survival, and ultrasound (USG) is commonly utilized for its safety and efficiency, especially in women with dense breast tissue. However, manual interpretation of USG images is time-consuming and subjective. This study presents a deep learning-based approach to automatically classify breast tumors into benign, malignant, or normal categories using USG images. Two public datasets, BUSI and Mendeley, were combined to create a dataset of 1,028 labeled images. Preprocessing steps included checking duplicate, normalization, noise reduction, contrast enhancement, and data augmentation. A total of nine models were evaluated, including six CNN-based architectures (MobileNetV2, MobileNetV3Large, VGG-16, ResNet-50, InceptionV3, EfficientNetB0) and three Transformer-based models (ViT, MobileViT, LeViT). Among all models, MobileViT achieved the highest performance with 93% accuracy, 95% precision, 91% recall, and 92% F1-score. This model outperformed ResNet-50, MobileNetV2, and MobileNetV3Large (each with 86% accuracy), as well as ViT (83%) and LeViT (75%). Grad-CAM visualizations confirmed MobileViT focus on tumor-relevant regions, enhancing interpretability. These results emphasize MobileViT strength in extracting both local and global features from images, making it well-suited for automated breast cancer detection. The integration of Explainable AI also enhances the model's transparency and relevance for clinical use.