MSCADMpox: A novel multi-stage classification model for effective monkeypox classification

Abstract Monkeypox (Mpox) is a skin disease that appeared after the COVID-19 pandemic and spread to 127 countries. Early diagnosis of Mpox according to symptoms is difficult because it is closer to measles and chickenpox symptoms, while both lesions are transmitted through respiratory droplets. However, according to skin imaging, computer-aided diagnosis systems (CADs) help diagnose the disease in the early stages. In this study, a novel model called multi-stage CAD for Monkeypox (MSCADMpox) is proposed. The proposed MSCADMpox comprises multiple stages. Firstly, a new proposed technique called RSWGAN-GP has been introduced to balance the dataset and address the overfitting issues. Secondly, features are extracted by using different handcrafted feature extraction techniques and four pre-trained models, such as Vision Transformer, VGG16, VGG19, and ResNet50. At the third stage, a new feature selection technique, MBGWO, was applied to select the most suitable features. MBGWO enhances optimal feature selection by identifying high-value frequent features and repeatedly refining the selection based on the best solution, mean, and standard deviation among all wolves. Finally, two datasets have been used, the Monkeypox Skin Lesion Dataset (MSLD) and the Monkeypox Skin Image Dataset (MSID), to ensure the effectiveness of the proposed model. To automate the diagnosis process, the proposed model uses a support vector machine (SVM) in the binary classification of the MSLD dataset and a random forest for the multiclassification in the MSID dataset that contains four skin lesions. To assess the effectiveness of the proposed model, comprehensive performance metrics such as accuracy, precision, recall, F1-score, and specificity are employed. The proposed model using the MSLD has achieved 90.54% accuracy, 91.43% precision, 88.49% recall, 90.14% F1-score, and 90.5% specificity. Moreover, the MSCADMpox has achieved 91.88% accuracy, 92.21% precision, 91.88% recall, 91.62% F1-score, and 96.93% specificity in the case of the MSID dataset.