Enhancing Skin Disease Diagnosis with a Dual Vision Transformer Model

An increasing number of people are experiencing skin problems, and it is never easy for a clinician to make a correct diagnosis. One potential solution to these problems is the use of deep learning for skin disease diagnosis. In this paper, a new kind of neural network is proposed for diagnosing skin diseases. We provide a new dual Vision Transformer (ViT) model that has two patches for image embedding and transformer encoding, as well as two multi-level permutation (MLP) heads to combine the two models, in light of the fact that the research datasets are images of skin diseases. To get deep characteristics from images, the proposed framework employs a ViT model that is both appropriate and effective. As a first step in pre-processing, images are first divided into uniformly sized patches, and then compressed into a string of tokens. The second step is to process the flattened patches using two transformer encoders that are connected in parallel. For the transformer encoder, there are $N_{i}, N_{j}$ layers, where each layer has two sublayers. The first component is a self-attention unit with multiple heads, while the second component is a feedforward network with all connections enabled. After normalizing the inputs to each of the two sublayers, we evaluate the strength of the residual connections between them. The implementation of a classification block follows the concatenation of MLP heads. The block is made up of two layers: one that is flattened and one that is dense and optimized for batch processing. The experimental results of the method show classification accuracies of 88% (ISIC2016), 90% (ISIC2017), 91% (HAM10000), and 88% (Skin_disease_v1). The proposed model represents an advancement over state-of-the-art methods and signifies progress in identifying skin diseases.