Hybrid Transformer Model for Skin Lesion Classification with Focuses on Monkeypox Diagnosis

Monkeypox is a serious health problem that needs immediate diagnosis and quarantine of infected individuals. Distinguishing Monkeypox from other cutaneous lesions, including Chickenpox or Measles, is important since misdiagnosis may have severe consequences. Visual differentiation among these lesions is frequently impractical. Conventional approaches such as dermoscopy or high-resolution ultrasound imaging, though helpful, are predicated on subjective analysis and take time. In this research, a combination of EfficientNetB3-Transformer-MLP is suggested for quantitative and objective classification of the six types of skin lesions: Chickenpox, Cowpox, HFMD, Healthy, Measles, and Monkeypox. The model combines EfficientNetB3, ImageNet pretrained as a feature extractor, with a Vision Transformer (ViT) to capture interpatch interactions and an MLP for ultimate classification, utilizing transfer learning and fine-tuning. The hybrid model contrasted with single EfficientNetB3 and Transformer models, along with other widely used pretrained deep learning networks, utilizing a transfer learning framework. The proposed hybrid model presented a better performance in terms of accuracy. The outcomes reflect competitive performance against existing studies, specifically in having balanced metrics against all classes, even with the drawbacks of class imbalance. The proposed model provides a trustworthy approach for researchers and dermatologists to speedily and effectively diagnose skin lesions, particularly in urgent situations such as Monkeypox outbreaks.