Monkeypox Diagnosis Using MRMR-Based Feature Selection and Hybrid Deep Learning Models: ResNet50V2, NASNetMobile, and InceptionV3

Monkeypox, like many other epidemics diseases, has been spreading rapidly. Its transmission through both respiratory droplets and physical contact has significantly contributed to its fast dissemination. The emergence of the first major outbreaks in the African region in 2022, followed by the disease spreading at an epidemic level, has raised global concerns. Although this potentially fatal disease can be partially detected through PCR methods, it often exhibits symptoms similar to other skin diseases, making accurate diagnosis challenging. At this point, computer-aided detection systems, particularly those based on image processing techniques, become crucial. The primary aim of this study is to enable the automatic diagnosis of monkeypox using deep learning methods by enhancing classification performance through the selection of the most significant features among multiple models. In this study, a hybrid deep learning approach is proposed that integrates transfer learning models such as ResNet50V2, NASNetMobile, and InceptionV3 with the mRMR (Minimum Redundancy Maximum Relevance) feature selection method. The features extracted from each model were concatenated to form a unified feature vector, from which the 10 most relevant features were selected using the mRMR algorithm. Finally, classification was performed based on these selected features. Experiments were conducted on three different datasets—MSLD, MSCI, and MSID—containing various skin lesion diseases. The proposed approach achieved accuracy rates of 92.00%, 92.50%, and 87.65%, respectively. Among these, the highest accuracy was observed on the MSCI dataset, with a rate of 92.50%. This hybrid approach demonstrated high performance across diverse datasets and significantly contributed to clinical diagnosis processes by enabling the accurate identification of not only monkeypox but also other visually similar skin lesions.