Evaluating the Effectiveness of Graph Neural Networks on an Augmented Dataset forMelanoma Skin Cancer Detection

Early detection of malignant melanoma is critical for patient survival, but even trained clinicians achieve only moderate accuracy with traditional visual inspection. Recent work demonstrates that deep convolutional neural networks (CNNs) can rival dermatologists in dermoscopy-based melanoma classification. Nevertheless, CNNs require large, balanced datasets, which are lacking for rare skin lesion classes. This study considered augmenting the ISIC 2019 skin lesion dataset with GAN-generated images to balance the benign and malignant classes and proposed a graph-based classification approach. The dermoscopic images are segmented into super pixels (using Simple Linear Iterative Clustering (SLIC)), a region-adjacency graph of super pixel nodes is constructed, and graph neural networks, including Graph Convolutional Network (GCN), Graph Attention Network (GAT), and a Dynamic Attention Variant graph, are trained. To mitigate the imbalance, we used focal loss and balanced resampling approaches. On the augmented dataset, GNN models achieve higher sensitivity and AUC than a CNN baseline (e.g., GAT ROC = 0.92 vs. CNN ROC = 0.87), indicating the benefit of graph-based relational features. Although based on other experimental results, all models tend to overfit to synthetic images, as performance drops when evaluated on a held-out set of real images. Additionally, training on a smaller, balanced, and real-only subset improves real-world performance and continues to favor GNNs over CNNs. Therefore, these findings suggest that graph representations can enhance melanoma detection in augmented data regimes but also highlight the pitfalls of synthetic augmentation and the need for robust evaluation.