A Skin Lesion Classification Approach Using Transfer Learning to Detect Monkeypox

Human monkeypox cases have been increasing in recent outbreaks, raising global concern. Historically, the virus was confined to certain regions of Africa. However, in recent times, cases have been found in areas with no prior history of monkeypox. The recent outbreak of monkeypox underscores the importance of utilizing deep learning-based approaches to aid in its diagnosis. In this study, we put forward an analysis of skin lesion classification that involves four classes: Monkeypox, Chickenpox, Measles, and Normal. We utilized various convolutional neural network (CNN) models pre-trained on ImageNet to extract features from lesion images, including InceptionV3, DenseNet, EfficientNet, MobileNet, ResNet50, ResNet101, ResNet152 and their hybrid models. The extracted high-level features were then classified using classifiers such as Logistic Regression and Multi-Layer Perceptron (MLP) models. Multiple experiments were conducted to determine the optimal combination of feature extractors and classifiers. Of all the combinations tested, the pairing of ResNet50 for feature extraction and Logistic Regression for classification delivered the best results, achieving an overall accuracy of 94.16%. The macro-average precision, recall, and F1-score were 92.31%, 93.77%, and 92.95%, respectively, while the weighted average precision, recall, and F1-score reached 94.43 %, 94.16%, and 94.21 %. These results demonstrate the effectiveness of transfer learning to extract relevant features and Logistic Regression for accurate classification in a multiclass skin lesion dataset, contributing to the advancement of automated disease detection methods.