Breast Cancer Classification Using Graph Learning Techniques

Accurate classification of breast cancer tumors is a critical factor in improving treatment outcomes, as distinguishing between benign and malignant tumors has a significant impact on treatment decisions. Despite the implementation of effective screening and treatment protocols, breast cancer continues to represent a significant cause of death among women on a global scale; however, the early detection of the disease has the potential to markedly reduce mortality rates. In this study, we employ graph learning models-specifically GraphSAGE and Auto-Encoders-on graph structures derived using Mutual k-Nearest Neighbors from the Diagnostic Wisconsin Breast Cancer dataset (WBCD) to enhance classification accuracy. We explore both inductive and transductive approaches to graph learning and assess model performance. By taking advantage of GraphSAGE's ability to understand and track both local and global structural information, we achieve superior predictive performance compared to traditional models. Our results, based on 10-fold cross-validation, demonstrate an accuracy of 98.07%, precision of 98.02%, recall of 98.01%, F1 score of 97.96%, and an AUROC of 99.54%. These findings suggest that graph-based learning methods significantly improve breast cancer classification, offering the potential for more effective diagnostic tools in clinical practice.