Softflatten-Net: A Deep Convolutional Neural Network Design for Monkeypox Classification From Digital Skin Lesions Images

In scenarios where smallpox has been eradicated, the most important human orthopoxvirus pandemic is monkeypox (MP), which is a zoonotic disease that spreads more often. MP is difficult to diagnose, as it has characteristics similar to smallpox, chickenpox (CP), and measles (MS). Polymerase chain reaction (PCR), computerized histopathological nucleic acid amplification test (NAAT), immunohistochemistry, enzyme-linked immunosorbent assay (ELISA), and electron microscope-based skin lesion testing are used for MP detection. But, these tests are time-consuming and may lead to false decision. Therefore, automation in MP detection using skin lesions images is needed for developing affordable and fast solutions. In this work, we have proposed a deep convolutional neural network architecture named the Softflatten network (Softflatten-Net) to identify and categorize MP using skin lesions. This architecture helps to avoid gradient vanishing and overfitting problems. Classification has been conducted in seven distinct scenarios: MP versus normal (NH), MP versus MS, MP versus CP, MP versus other, MP versus CP versus NH, MP versus CP versus MS, and MP versus CP versus MS versus NH. Among binary classification tasks, the Softflatten-Net achieved the maximum classification accuracy (Ac) of 97.25%, a sensitivity (Sen) of 92.94%, and a precision (Pr) of 96.48% in classifying MP from other categories of skin images. In the three-class classification task (MP versus CP versus NH), the proposed network achieved an Ac of 94.71%, a Pr of 92.11%, and a Sen of 87.87%. In the four-class classification scenario, namely, MP versus CP versus MS versus NH, the proposed model achieved the Ac, Pr, and Sen values of 90.73%, 84.21%, and 69.38%, respectively. Furthermore, we have considered 4241 skin images from IEEE data port for nine-class classification and 28771 images from Dermnet for 20-class classification to validate the performance of the proposed Softflatten-Net architecture. The proposed architecture performed satisfactorily on all the three datasets considered in this work and outperformed the existing state-of-the-art models in the literature. A postimage analysis is performed via local interpretable model-agnostic explanations (LIMEs), gradient class activation maps (G-CAMs), and occlusion sensitivity (OS) approaches. In addition, to analyze misclassification, the stochastic neighbor embedding distributed with t (t-SNE) is employed. Visualization techniques aided in identifying the image regions most pertinent for classification. The proposed network can be utilized for early and reliable detection of MP using images of skin lesions.