Context-Aware Graph Neural Network for Skin Lesion Classification

The accurate classification of skin lesions, particularly melanoma, is vital for the early detection and effective treatment of skin cancer. Although deep learning models such as convolutional neural networks (CNNs) have achieved remarkable success in dermoscopic image analysis, they often overlook valuable structured metadata (e.g., patient demographics, lesion location, and type) that provide essential diagnostic context. We present a graph-driven multimodal framework that jointly models visual and metadata information for skin lesion classification. Our approach uses a frozen CNN backbone to extract deep visual representations via dual pooling (mean and max), which are concatenated with encoded metadata and partitioned into subspaces. These subspaces are treated as nodes within a graph, where a graph neural network (GNN) captures intra-sample dependencies between feature subspaces and clinical attributes to refine lesion representations. Experiments on four public benchmarks: ISIC2024, HAM10000, PAD-UFES-20, and HIBA, demonstrating consistent performance gains over several state-of-the-art (SOTA) approaches, with relative accuracy improvements ranging from +0.5% to +8.8% in datasets. The results highlight the potential of graph-based modeling of metadata and image features to build more robust and clinically informed skin cancer classifiers.