Early detection and accurate classification of skin diseases is clinically important and required for proper diagnosis and treatment because delayed diagnosis can lead to inadequate treatment and increased patient risk, especially when visually similar conditions require different management approaches. However, automated skin disease detection remains challenging due to subtle inter-class differences and the need to capture both individual lesion textures and patterns. In this paper, we propose a lightweight hybrid model to detect and classify skin diseases accurately with less computational complexity. The proposed hybrid deep learning architecture combines complementary representations from two models, MaxViT and MobileNet. MaxViT uses a multi-axis self-attention mechanism that decomposes global attention into block-wise and grid-wise operations, reducing the quadratic self-attention complexity, while MobileNet effectively extracts spatial features using depth-wise separable convolutions with minimal parameters. We evaluate the performance of the proposed hybrid method on two different Datasets, including eight classes of skin disease images. The proposed hybrid model obtains an accuracy of 99.57% and 98.67% using Dataset 1 and Dataset 2, respectively. The results indicate that by combining complementary features from two models can ensure better screening accuracy while maintaining less computational complexity and processing time, which is required for real-time clinical workflow decision support.
Ahmed et al. (Sun,) studied this question.