Leveraging Edge Optimize Vision Transformer for Monkeypox Lesion Diagnosis on Mobile Devices

Abstract

Rapid and precise diagnostic tools for Monkeypox (Mpox) lesions are crucial for effective treatment because their symptoms are similar to those of other pox-related illnesses, like smallpox and chickenpox. The morphological similarities between smallpox, chickenpox, and monkeypox, particularly in how they appear as rashes and skin lesions, which can sometimes make diagnosis challenging. Chickenpox lesions appear in many simultaneous phases and are more diffuse, often beginning on the trunk. In contrast, monkeypox lesions emerge progressively and are typically centralized on the face, palms, and soles. To provide accessible diagnostics, this study introduces a novel method for automated monkeypox lesion classification using the HMTNet (Hybrid Mobile Transformer Network). The convolutional layers and Vision Transformers (ViT) are combined to enhance the spatial features. In addition, we replace the classical MHSA (Multi-head self-attention) with the WMHSA (Window-based Multi-Head Self-Attention) to effectively capture long-range dependencies within image patches and depth-wise separable convolutions for local feature extraction. We trained and validated HMTNet on the two datasets for binary and multiclass classification. The model achieved 98.38% accuracy for multiclass classification using cross-validation and 99.25% accuracy for binary classification. These findings show that the model has the potential to be a useful diagnostic tool for monkeypox, especially in environments with limited resources.