Classification of monkeypox lesions using a deep learning-based hybrid model

In this study, a deep learning-based hybrid model was developed to support the early and accurate diagnosis of monkeypox, a global health concern, through skin lesions. The proposed model utilizes features automatically extracted from four distinct pre-trained deep learning architectures (InceptionV3, SqueezeNet, VGG16, and VGG19), which were subsequently optimized using seven different filter-based feature selection methods (Information Gain, Gain Ratio, Gini Index, ReliefF, FCBF, ANOVA, and Chi2). The resulting feature sets were classified using traditional machine learning algorithms, such as k-Nearest Neighbors, Logistic Regression, and Artificial Neural Networks (ANN). Experimental results demonstrate that the hybrid model—wherein deep features are optimized via feature selection methods and classified through machine learning algorithms—yields superior classification performance compared to standard transfer learning-based classification alone. Notably, a maximum accuracy of 0.9510 was achieved by employing Information Gain and Gain Ratio feature selection methods at a 95% feature retention rate, coupled with the ANN classifier. Furthermore, it was determined that the application of feature selection leads to limited but consistent improvements in performance compared to non-filtered scenarios, effectively reducing redundant noise while preserving information density. The findings provide competitive and higher performance values relative to similar monkeypox skin lesion classification studies reported in the literature. In conclusion, this study demonstrates that the integrated use of transfer learning, feature selection, and traditional machine learning algorithms—without the need for complex architectures—offers a feasible, reproducible, and high-performance solution for medical image analysis.