Evaluating Study Between Vision Transformers and Pre-trained CNN Learning Algorithms to Classify Breast Cancer Histopathological Images

Breast cancer exhibits a higher incidence rate among women compared to other forms of cancer. In the Kingdom of Saudi Arabia, there exists a higher propensity for the development of breast cancer in women aged 40 and above, as compared to their younger counterparts. Convolutional neural networks (CNNs) are the well-known approach for classifying breast cancer mammographic images. The utilization of convolutional neural networks has proven to be effective in the interpretation of mammograms and the study of medical imaging recent times, a new revelation called Vision Transformers (ViT) has been able to attend to global information in earlier layers, facilitating a more adaptable methodology for the exploration and identification of features. This study conducted a comparative analysis between the vision transformer ViT_16 and the ResNet18 convolutional neural network, which is considered the current leading model in the field. The models were trained using datasets obtained from the King Abdulaziz University Breast Cancer Mammography Dataset (KAU-BCMD) version 1. The results have been identified by evaluating the accuracy metrics. ResNet18 outperformed ViT_16 by 8% in overall accuracy. However, ViT_16 exhibit superior performance in detecting suspicious abnormality. This emphasizes the potential benefit of ViT_16 model in classifying breast cancer images. Overall, ViT_16 model performance could be improved when trained with sufficient data.