Attention to Monkeypox: An Interpretable Monkeypox Detection Technique Using Attention Mechanism

Abstract

In the wake of COVID-19, rising monkeypox cases pose a potential pandemic threat. While less severe than COVID-19, its increasing spread underscores the urgency of early detection and isolation to control the disease. The main difficulty in diagnosing monkeypox arises from its prolonged diagnostic process and symptoms that are similar to those of other skin diseases, making early detection and isolation challenging. To address this, the deployment of deep learning models on edge devices presents a viable solution for the rapid and accurate detection of monkeypox. However, the resource constraints of edge devices require the use of lightweight deep learning models. The limitation of these models often involves a trade-off with accuracy, which is unacceptable in the context of medical diagnostics. Therefore, the development of optimized deep learning models that are both resource-efficient for edge computing and highly accurate becomes imperative. To this end, an attention-based MobileNetV2 model for monkeypox detection, capitalizing on the inherent lightweight design of MobileNetV2 for effective deployment on edge devices, is proposed. This model, enhanced with both spatial and channel attention mechanisms, is tailored for rapid and early-stage diagnosis of monkeypox with better accuracy. We significantly improved the Monkeypox Skin Images Dataset (MSID) by incorporating a broader range of classes for similar skin diseases, thereby substantially enriching and diversifying the training dataset. This helps better distinguish monkeypox from other similar skin diseases, particularly in its early stages or when a detailed medical examination is unavailable. To ensure transparency and interpretability, we incorporated Gradient-weighted Class Activation Mapping (Grad-CAM) and Local Interpretable Model-Agnostic Explanations (LIME) to provide clear insights into the model's diagnostic reasoning. Finally, to comprehensively assess the performance of our model, we employed a range of evaluation metrics, including Cohen's Kappa, Matthews Correlation Coefficient, and Youden's J Index, alongside traditional measures like accuracy, F1-score, precision, recall, sensitivity, and specificity. The attention-based MobileNetV2 model demonstrated impressive results, outperforming the baseline models by achieving 92.28% accuracy in the extended MSID dataset, 98.19% in the original MSID dataset, and 93.33% in the Monkeypox Skin Lesion Dataset (MSLD) dataset.