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

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Pathology

Background:

• Breast cancer image classification is complex due to high-resolution pathological images and intricate feature distributions.
• Graph neural networks (GNNs) capture local structures but struggle with generalization and shortcut features.

Purpose of the Study:

• To develop a novel causal discovery attention-based graph neural network (CDA-GNN) for enhanced breast cancer image classification.
• To improve model interpretability and robustness by disentangling causal and shortcut features.

Main Methods:

• Converted high-resolution images to graph data via superpixel segmentation.
• Employed a causal attention mechanism to identify key causal features.
• Utilized a backdoor adjustment strategy to separate causal from shortcut features.

Main Results:

• Achieved 86.36% classification accuracy on the 2018 BACH breast cancer dataset.
• Demonstrated superior performance and generalization through F1-score and ROC metrics.
• Validated the model's interpretability and robustness.

Conclusions:

• The CDA-GNN model offers powerful automated cancer image analysis capabilities.
• It enhances clinical applications by improving diagnostic efficiency and accuracy.
• The model aids in early breast cancer detection and reduces healthcare professional workload.