Truncated MobileNetV2 Sparse Vision Graph Attention Model for Explainable Monkeypox Disease Classification

Abstract

The current outbreak of monkeypox (mpox) presents challenges for timely and accurate diagnosis due to the disease’s diverse and unusual skin lesion patterns. Traditional deep learning models, such as Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs), struggle with these irregular features because they rely on rigid, grid-based methods. To address this, we introduce the Truncated MobileNetV2 Sparse Vision Graph Attention (TMSVGA) model. TMSVGA combines components of MobileNetV2, which focuses on identifying smaller details, with a Sparse Vision Graph Attention block enhanced by a Squeeze-and-Excitation (SE) mechanism to improve channel-wise attention. This approach enhances the understanding of complex and long-distance relationships, emphasizing diagnostically significant regions and improving classification precision. We optimized TMSVGA using the Optuna framework for automated hyperparameter tuning. Additionally, Gradient-weighted Class Activation Mapping (Grad-CAM) and Local Interpretable Model-Agnostic Explanations (LIME) provided interpretable visualizations, highlighting influential regions in decision-making. The TMSVGA model was validated on the Monkeypox Skin Images Dataset (MSID), achieving 96.79 % accuracy, 96.90 % precision, 95.34 % recall, 96.08 % F1-score, and 95.37% Matthews Correlation Coefficient (MCC). These results demonstrate that TMSVGA outperforms existing models, particularly in handling irregular lesion patterns. By achieving high diagnostic accuracy and precision, our study showcases the potential of Vision Graph Neural Networks (ViGNNs) in advancing medical image analysis for diseases with non-uniform spatial patterns. Furthermore, the lightweight architecture of TMSVGA ensures suitability for mobile and resource-constrained diagnostic applications.