Improving mpox identification under imbalanced data using geometric transformation and color augmentation with graph neural networks
Mpox lesion identification from skin images remains challenging because lesion appearances often overlap with other dermatological diseases and available data are commonly imbalanced. This study proposes a graph-based multi-class classification framework that combines class-balancing augmentation, convolutional neural network (CNN)-based feature extraction, and graph neural classification. Experiments were conducted on the Mpox Skin Lesion Dataset Version 2.0 using three data configurations: (i) the original imbalanced data, (ii) a balanced version generated through geometric transformation augmentation, and (iii) a balanced version generated through combined geometric transformation and color-space augmentation. Image features were extracted using eight CNN backbones (EfficientNetB4, AlexNet, VGG16, ResNet50, DenseNet121, GoogleNetV3/InceptionV3, MobileNetV2, and LeNet), then transformed into graphs by selecting k-nearest-neighbor candidates and retaining edges according to cosine-similarity filtering, and finally classified using graph convolutional networks (GCN) and graph attention networks (GAT). Across 48 model combinations, data balancing improved macro-level performance compared with the original imbalanced setting. The best overall result was achieved by GCN with VGG16 on the geometrically and color-augmented balanced data, reaching 92.00% accuracy, 91.55% macro F1-score, 98.81% macro area under the receiver operating characteristic curve (one-versus-rest), 93.54% macro precision, and 90.32% macro recall. The best GAT result was obtained by MobileNetV2 on the geometrically augmented balanced data, with 90.67% accuracy, 89.25% macro F1-score, 98.48% macro area under the receiver operating characteristic curve (one-versus-rest), 89.28% macro precision, and 90.27% macro recall. Although GCN showed a more consistent empirical performance trend than GAT, the difference was not statistically significant. These findings indicate that integrating balanced augmentation, CNN-derived representations, and graph learning is a promising strategy for mpox lesion identification within the scope of the present experiments; however, broader external validation and explainability-oriented analysis are still required before real-world clinical deployment.