DermVit: A Standalone Transformer-Based System for Multi-Class Skin Disease Detection and Early Diagnosis

With the increasing demand for early and accurate diagnosis in healthcare, the detection and classification of skin diseases remain a critical challenge. Conventional convolutional neural networks (CNNs) often struggle to capture the fine-grained features necessary for distinguishing between various skin conditions, leading to limitations in automated diagnosis. To address this issue, this paper presents DermViT, a Vision Transformer (ViT)-based deep learning system designed for multi-class skin disease detection. Instead of relying solely on CNN architectures, the proposed approach leverages the pre-trained google/vit-base-patch16-224 model, fine-tuned with dynamic augmentation and feature extraction techniques to enhance performance. To improve model robustness, several strategies were employed, including data augmentation, class weighting, and advanced image preprocessing using CLAHE and Gamma Correction. Dataset balancing was achieved through oversampling, undersampling, and augmentation techniques such as Cutout, CutMix, Mixup, and RandAugment. Additionally, a layer-wise freezing strategy was applied to determine the optimal ViT configuration, while a customized dropout layer was incorporated to mitigate overfitting. The system underwent rigorous evaluation through module, integration, and performance testing, achieving an overall accuracy of 94.48% on test data. Experimental results demonstrate that the ViT-based model delivers high precision, recall, and F1 scores across all four skin disease classes. These findings highlight DermViT as a promising solution for automated skin disease detection, with potential applications in clinical diagnostics and telemedicine.