Breast Cancer Detection Based on Graph Neural Networks

Early detection of breast cancer is crucial for improving patient survival rates. However, existing automated detection methods based on medical imaging (such as mammography, ultrasound, MRI) or pathological slides often suffer from high false positive/negative rates, limited sensitivity to subtle structural changes, and heavy reliance on expert experience. This study introduces an innovative graph node anomaly detection framework for breast cancer diagnosis. The core concept involves modeling medical data as a graph structure where nodes represent cases and edges capture spatial proximity or feature similarity between them. We identify potential cases as "abnormal nodes" within the graph. To tackle this challenge, we propose the Breast Cancer Graph Anomaly Detection framework, which employs an enhanced graph neural network (GNN) architecture to reconstruct both node attributes and topological structure through self-supervised learning. The core innovation lies in leveraging GNN's powerful relational reasoning capabilities to learn graph pattern representations from "normal" cases. A key breakthrough is our newly designed anomaly scoring mechanism that jointly optimizes reconstruction errors for both node attributes and graph structure. This mechanism effectively identifies nodes exhibiting significant deviations from normal patterns in feature expression or topological connectivity. By aggregating node-level anomaly scores, the system achieves comprehensive classification of malignancy status for entire images or cases. We conducted extensive experimental evaluations on the publicly available breast cancer dataset WBCD. The results demonstrate that our proposed method outperforms traditional machine learning approaches, mainstream deep learning image classification models (e.g., ResNet, DenseNet), and other graph learning methods in key metrics including time efficiency and AUC (area under the curve). The method also exhibits excellent interpretability, providing clinicians with valuable decision-support information. This study not only demonstrates the great potential of graph node anomaly detection in computer-aided diagnosis (CAD) of breast cancer, but also provides a new idea for using graph structure learning to model complex medical data relationships and detect local anomaly patterns, which has important theoretical significance and clinical application value.