A comparative study of advanced deep learning architectures for breast cancer classification on ultrasound and histological images

Breast cancer persists as a critical global health issue and is the most frequently diagnosed malignancy worldwide. Early and accurate detection is paramount, as it significantly improves survival rates and treatment efficacy. While medical imaging, particularly ultrasound and histopathology, is central to diagnosis, deep learning has emerged as a promising paradigm to enhance diagnostic accuracy and support clinical workflows via automated classification. This study provides a comprehensive, systematic comparison of state-of-the-art deep learning architectures using two standardized public datasets: Breast Ultrasound Images (BUSI) and Breast Cancer Histology (BACH). Nearly 30 Convolutional Neural Network (CNN) and Vision Transformer (ViT) models were rigorously assessed, incorporating transfer learning and data augmentation to mitigate challenges of class imbalance and limited data. Our findings reveal that optimal model performance is strongly modality-dependent. For the BUSI dataset, CNN-based architectures demonstrated superior performance, with EfficientNetV2-Small achieving a peak accuracy of 90.52 %. Conversely, on the BACH dataset, the multi-scale ViT (MViTv2-Base) and the lightweight CNN (MobileNetV3-Large-100) jointly achieved the highest accuracy of 91.67 %. These results establish strong performance benchmarks and offer evidence-based guidance for selecting architectures tailored to specific imaging modalities. By identifying computationally efficient yet high-performing models, this work contributes to the advancement of deployable computer-aided diagnosis systems for diverse clinical contexts, with the ultimate goal of enhancing patient care and outcomes.