An explainable AI for breast cancer classification using vision Transformer (ViT)

Abstract

Manual classification of breast cancer (BC) through an optical microscope is regarded as an essential task throughout clinical routines, necessitating highly skilled pathologists. Computer-aided diagnosis (CAD) techniques based on deep learning (DL) are developed to assist the pathologists in making diagnostic decisions. Nevertheless, the black-box nature and the absence of interpretability and transparency of these DL-based models render their application highly difficult in sensitive and critical medical applications. In addition to providing explanations for the model predictions, explainable artificial intelligence (XAI) strategies help to gain the trust of clinicians. The current Convolutional Neural Network (CNN) architectures have limitations in capturing the global feature information details present in BC histopathological images. To overcome the challenge of long-range dependenciesin CNN-based models, Vision Transformer (ViT) architectures have recently been created.

These architectures have a self-attention mechanism that enables the analysis of images. As a result, the network is able to record the deep long-range dependence between pixels. The present work aims to develop an effective CAD tool for BC classification. In this study, we investigated a deep ViT architecture trained to perform binary lesions classification (malignant versus benign) using histopathology images. Various XAI techniques have been implemented: Gradient-Weighted Class Activation Mapping (Grad-CAM), Vanilla gradient, Integrated gradients, Saliency Maps, Local Interpretable Model Agnostic Explanation (LIME), and Attention Maps to highlight the most important features of the model prediction outcomes. The evaluation task was performed using the publicly accessible benchmark dataset BreakHis. Based on the research outcomes, our suggested ViT architecture demonstrates competitive performance, surpassing state-of-the-art CNN models in the analysis of histopathological images. Furthermore, the proposed models provide precise and accurate interpretations, reinforcing their reliability. Therefore, we can affirm that the proposed CAD system can be effectively integrated into clinical diagnostic routines, offering enhanced support for medical professionals.