Hybrid CNN-Transformer Architecture With Xception-Based Feature Enhancement for Accurate Breast Cancer Classification

Abstract

Breast cancer is one of the most common causes of death in women worldwide. While accurate diagnosis it from histopathological images is vital, the process often relies heavily on pathologists' skills, leading to inconsistent results and lengthy evaluation times. To address this, computer-aided diagnostic (CAD) techniques are advised. This study introduces a deep learning (DL) approach that integrates Xception and Transformer architectures to improve breast cancer classification from histopathological images. The proposed model leverages Xception for local feature extraction, while a Transformer captures global contextual features, thereby overcoming the limitations of conventional models in handling both local and global dependencies in medical images. The architecture is evaluated on two publicly available datasets, BreaKHis and IDC. Our proposed model achieved accuracy ranging from 96.15% to 100% in the magnification-dependent (MD) scenario, from 94.82% to 99.62% in the magnification-independent (MI) scenario on the BreaKHis dataset and 91% in the binary classification of the IDC dataset. This approach surpasses existing models in both binary and eight-class classification. This can reduce the diagnostic workload, decrease diagnostic variability and provide rapid, reliable support for clinical decision-making.