Breast cancer image classification based on H&E staining using a causal attention graph neural network model.

Abstract

Breast cancer image classification remains a challenging task due to the high-resolution nature of pathological images and their complex feature distributions. Graph neural networks (GNNs) offer promising capabilities to capture local structural information but often suffer from limited generalization and reliance on shortcut features. This study proposes a novel causal discovery attention-based graph neural network (CDA-GNN) model. The model converts high-resolution image data into graph data using superpixel segmentation and employs a causal attention mechanism to identify and utilize key causal features. A backdoor adjustment strategy further disentangles causal features from shortcut features, enhancing model interpretability and robustness. Experimental evaluations on the 2018 BACH breast cancer image dataset demonstrate that CDA-GNN achieves a classification accuracy of 86.36%. Additional metrics, including F1-score and ROC, validate the superior performance and generalization of the proposed approach. The CDA-GNN model, with its powerful automated cancer image analysis capabilities and strong interpretability, provides an effective tool for clinical applications. It significantly reduces the workload of healthcare professionals while facilitating the early detection and diagnosis of breast cancer, thereby improving diagnostic efficiency and accuracy.