BreAST-Net: Breast Cancer Classification Using an Advanced Swin Transformer on Histopathological Images

Breast cancer ranks among the foremost causes of mortality among women in Indonesia, with both incidence and mortality rates rising annually.Timely detection is essential for identifying suitable treatments according to the specific type and subtype of cancer, hence decreasing associated mortality rates.Diverse deep learning methodologies, notably Convolutional Neural Networks (CNNs), have been employed for the medical image classification.Nonetheless, CNNs possess limitations due to their localized feature extraction, which hinders their ability to capture global interactions among cells in histopathological images.This study proposes the Breast Advanced Swin Transformer Network (BreAST-Net), a model based on the Swin Transformer enhanced by adjustments to the global activation function, substituting the standard Gaussian Error Linear Unit (GELU) activation function.In experiments utilizing the BreakHis dataset, BreAST-Net demonstrated improved performance relative to the CNN-based models and conventional Swin Transformer, reaching the highest accuracy, precision, recall, and F1-score of 99.60%, 99.50%, 99.60%, 99.55% for breast cancer types and 97.57%, 97.04%, 97.57%, 97.24% for breast subtypes classification, respectively.The results suggest that BreAST-Net may be integrated into clinical diagnostic systems, offering an accurate and efficient method for the early breast cancer diagnosis.