<p>Medical image classification in federated healthcare environments is challenged by the difficulty of learning robust and generalized representations from heterogeneous, non-IID data distributed across multiple institutions. Conventional federated learning approaches primarily rely on visual features and often fail to capture structural relationships among medical images, leading to reduced classification performance and limited generalization across diverse clinical settings. This study proposes GraphMedFL, a federated learning framework that integrates knowledge graphs and Node2Vec-based structural embeddings to enhance classification performance in multi-source environments. Local knowledge graphs are constructed from image features to capture inter-sample relationships, and Node2Vec embeddings are fused with feature representations to enrich local model training. A similarity-aware adaptive aggregation strategy is introduced to address non-IID data distributions across clients. Experiments conducted on three breast cancer imaging datasets (BreakHis, CBIS-DDSM, and INbreast) demonstrate that GraphMedFL achieves superior performance compared to locally trained models, with an accuracy of 98.3% and strong precision-recall balance. The proposed GraphMedFL achieves an accuracy of 98.3% and consistently outperforms baseline federated learning and local training approaches across BreakHis, CBIS-DDSM, and INbreast datasets. It improves precision, recall, and F1-score by a measurable margin compared to conventional methods, demonstrating enhanced robustness and generalization in heterogeneous medical imaging environments.</p>