A Deep Learning Approach to Multi-Class Skin Disease Identification

Skin diseases affect more than one-third of the global population and are often underestimated in terms of their severity and consequences. Traditional diagnostic methods, such as visual inspection and biopsy, are time-consuming and dependent on the expertise of dermatologists, limiting their efficiency and scalability. In order to address these limitations, this work proposes the development of a multi-level, multi-class deep learning model for the classification of skin diseases, that can aid the dermatologist. The dataset comprises 31 classes with a total of 8,867 images sourced from databases such as HAM10000, ISIC Archive, Arsenic Healthy Skin, and DERMNET. After applying preprocessing and data augmentation, the dataset was expanded to 12,221 images to enhance uniformity and quality. The classification framework is structured into four hierarchical levels, with each hierarchical level using a different deep learning model designed uniquely to its specific classification task. The hierarchical structure enhances the model’s ability to detect slight variations between related classes of skin diseases thereby maximizing classification accuracy and generalization. The proposed multi-level model yielded a combined accuracy of $91.20 \%$, with a specificity of $95.36 \%$ and sensitivity of $92.86 \%$, demonstrating its effectiveness in accurately classifying skin diseases. This study also leverages the gradient-weighted class activation mapping (Grad-CAM) technique to identify and to visualize the relevant regions of interest, thereby the proposed system provides a comprehensive evaluation framework.