Fast Vision Transformer with Hierarchical Attention for Multi-Class Breast Cancer Classification in Histopathology Images

Abstract

Breast cancer (BC) is a leading cause of female mortality across the globe. The histopathology analysis of breast tissues is a more crucial tool for diagnosing and staging BC. In recent times, there has been a considerable increase in research examining the usage of deep-learning (DL) models for BC detection from histopathology images (HIs). DL is developed as a precious tool for BC detection by automatically learning discriminative features from medical images. Its capability to model complex patterns has significantly enhanced diagnostic accuracy and maintained earlier and more reliable cancer identification. Consequently, explainable artificial intelligence (XAI) methods have been integrated to show how DL models arrive at their conclusions. Therefore, this study presents a Transformer with Hierarchical Attention for Explainable Multi-Class Breast Cancer Detection (THAXMCBCD) model. The primary intention of this paper is to develop an intelligent system capable of accurately detecting and classifying BC from histopathology images. The proposed THA-XMCBCD model begins with a robust preprocessing stage, including color normalization, image enhancement, and pixel normalization to prepare the input image data for analysis. Following that, a fast vision transformer with hierarchical attention is utilized for feature extraction, which captures local cellular patterns and global tissue context. The extracted features are further passed into a heterogeneous graph attention network for BC classification, including Adenosis, Ductal Carcinoma, Fibroadenoma, Lobular Carcinoma, Mucinous Carcinoma, Papillary Carcinoma, Phyllodes Tumor, and Tubular Adenoma. Furthermore, the model is trained using the AdamW optimizer to ensure stable and efficient convergence. To provide interpretability, ScoreCAM is employed to visualize the region’s most influential in the classification decision. The THA-XMCBCD technique is evaluated using the benchmark BreakHis dataset under magnification factors of 100X and 400X. An extensive simulation analysis is performed, and the obtained results demonstrate the betterment of the THA-XMCBCD technique over the recent methods.