A neutrosophic set-based hybrid Swin transformer and graph neural network model for monkeypox diagnosis

Abstract

Monkeypox is a zoonotic viral disease caused by the Monkeypox virus (MPXV) of the Orthopoxvirus genus. Since infection primarily manifests as viral dermatological lesions, visual assessment is a crucial component of clinical evaluation. Therefore, the development of image processing-based machine learning systems for the automated detection of monkeypox-related skin lesions is highly important, not only for medical image analysis but also for computational approaches to viral disease classification. In this study, a hybrid deep learning model, Neutro-Swin-GNN, enhanced with the Neutrosophic Set (NS) approach, is proposed to distinguish monkeypox from similar skin lesions (chickenpox, measles, and normal skin). The model addresses the challenges of low contrast and visual similarity in medical images by combining the uncertainty modeling capability of NS with the hierarchical feature extraction capability of the Swin Transformer and the relational feature learning capability of Graph Neural Networks (GNNs). Experimental studies were conducted on the Monkeypox Skin Image Dataset (MSID). Model selection was performed using stratified fivefold cross-validation, which yielded an accuracy of 98.77 +/- 0.36%, and the final evaluation on an independent test set achieved an accuracy of 99.14%. The reported accuracy of 99.14% corresponds to the performance on the independent test set. The macro-average and weighted-average F1-scores were 0.99 and 0.98, respectively. In the comparative analyses, the proposed model was evaluated against MobileNetV2, ResNet50, Swin-V2, MobileNetV3, EfficientNetV2, and ShuffleNetV2 architectures, demonstrating competitive performance on the MSID dataset. This study presents an effective framework for monkeypox classification by integrating NS-based uncertainty modeling with a Swin Transformer-GNN hybrid framework.