Early detection of human Mpox: A comparative study by using machine learning and deep learning models with ensemble approach

Abstract

Objective: This study aims to enhance the early diagnosis of Mpox through machine learning (ML) and deep learning (DL) models, integrating an ensemble approach to improve classification accuracy.

Methods: We used the Mpox Skin Lesion Dataset v2.0, comprising six skin lesion categories, including chickenpox, cowpox, Mpox, measles, hand-foot-mouth disease, and healthy skin. Four models-Logistic Regression, K-Nearest Neighbors, Vision Transformer (ViT), and ConvMixer-were evaluated based on their classification performance. An ensemble model combining ViT and ConvMixer predictions was developed to further improve accuracy and robustness. Performance metrics such as accuracy, precision, recall, F1-score, and AUC were used for evaluation.

Results: The ViT model outperformed traditional ML models, achieving 93.03% accuracy in detecting Mpox lesions. The ensemble model further improved diagnostic performance, yielding balanced precision and recall across all lesion categories. The proposed approach demonstrated superior classification accuracy compared to previous studies, highlighting the efficacy of DL-based models in distinguishing Mpox from visually similar conditions.

Conclusion: The integration of ML and DL models in an ensemble framework significantly enhances Mpox detection. This AI-driven diagnostic approach offers a scalable, accurate, and efficient solution, particularly in resource-limited settings. Future research will focus on improving model interpretability, federated learning integration, and validation with real-world clinical data.

Keywords: Mpox; deep learning models; ensemble approach; machine learning models; six-class classification.

Figure 1.

Workflow diagram for skin lesion classification and detection.

Figure 2.

Receiver operating characteristic (ROC) curves of (a) LR, (b) KNN, (c) ViT, and (d) ConVmixer.

Figure 3.

Precision recall curve of (a) LR, (b) KNN, (c) ViT, and (d) ConvMixer.

Figure 4.

Confusion matrix of (a) LR, (b) KNN, (c) ViT, and (d) ConvMixer.

Figure 5.

Training and testing accuracy and loss curves for (a) the ViT model and (b) the ConvMixer model.

Figure 6.

Model outputs for studied DL models.

Figure 7.

Accuracy and loss curve of ensemble model.

Figure 8.

Confusion matrix of ensemble model.

Figure 9.

(a) ROC of ensemble model, (b) P-R curve of ensemble model.