Mammography Breast Cancer Classification Using Vision Transformers

Breast cancer is a worldwide concern, emphasizing the importance of early diagnosis for effective treatment. Convolutional networks (CNN) have long dominated medical vision tasks. Recently, vision transformers (VIT) have sparked interest in computer-aided diagnosis (CAD). Nonetheless, when ViTs are trained from scratch, they exhibit poorer performance than (CNNs), particularly in low-data scenarios. This discrepancy is even more pronounced in tasks involving medical image datasets, where data scarcity is a significant challenge. To this end, we proposes a Vision Transformer (ViT) model that has been pre-trained on extensive datasets, including (ImageNet) and is applied directly to sequences of mammography breast image patches for the classification of benign and cancerous tissue. The model was thoroughly assessed using the DDSM dataset containing 5970 benign and 7158 malignant Masses, yielding remarkable results, with an accuracy of 99.96%, precision of 99.95%, Sensitivity of 99.96%, F1 score of 99.96% for breast mass classification. Our study illustrates how trained attention mechanisms can be effectively utilized to develop high-performing computer-aided systems for breast cancer diagnosis. Importantly, this approach achieves these results while also minimizing computational resource requirements and reducing prediction time.