Contrast-Enhanced Spectral Mammography (CESM) has emerged as a promising technique for breast cancer detection, providing functional information on lesion vascularity through iodinated contrast enhancement. Despite its diagnostic potential, the automated analysis of CESM images remains relatively underexplored, with most existing approaches relying on Convolutional Neural Networks that process images as regular Euclidean grids and do not explicitly model the spatial relationships between tissue regions. In this study, we propose a graph-based framework for three-class classification of CESM subtracted images into Normal, Benign, and Malignant categories. The images are segmented into compact superpixel regions using the Simple Linear Iterative Clustering (SLIC) algorithm and represented as undirected graphs, where nodes encode 30-dimensional handcrafted radiomics features capturing intensity, texture, and shape characteristics. The spatial relationships between superpixels are modeled through k-Nearest Neighbor connectivity based on centroid proximity. Three Graph Neural Network (GNN) architectures are evaluated: Graph Convolutional Network (GCN), Graph Attention Network (GAT), and Graph Isomorphism Network (GIN). Across all configurations, GCN with $k=4$ achieves the best performance, with an AUC-ROC of 0.729, AUPRC of 0.553, and macro-F1 of 0.506, outperforming a ResNet-18 baseline trained on the same preprocessed images. These results highlight the potential of graph-structured representation for CESM-based breast cancer classification.