DermAI: A Vision Transformer-Based Web System for Multi-Class Skin Disease Detection

Abstract: Dermatological conditions represent a significant global health burden, often requiring specialized expertise for accurate diagnosis. Early detection of skin diseases, particularly malignancies like melanoma, is critical for improving patient outcomes. This paper presents DermAI, an advanced, production-ready skin disease detection system powered by a fine-tuned DINOv2-Base Vision Transformer (ViT) architecture. By leveraging self-supervised features distilled from large-scale image data, DermAI achieves a validation accuracy of 95.57% across 31 distinct skin disease classes. The proposed system integrates a high-performance deep learning backend with a user-friendly web interface, facilitating real-time inference and ranking predictions by confidence. Our results demonstrate that foundational vision models can be effectively repurposed for specialized medical diagnostic tasks with high precision and reliability. The global prevalence of dermatological conditions, coupled with a shortage of specialized dermatologists, necessitates the development of accessible, highly accurate automated diagnostic tools. This paper presents DermAI, an end-to-end web- based system designed for the classification of 31 distinct skin diseases. By integrating a fine-tuned DINOv2-Base Vision Transformer (ViT-B/14) with a robust full-stack web architecture (Python, Flask, HTML, CSS), the system delivers real- time, confidence-ranked predictions to the end-user. The underlying deep learning model -images, achieving a remarkable validation accuracy of 95.57% over 10 epochs. Through detailed case studies—including the high-confidence identification of malignant melanoma and benign fungal infections—this paper demonstrates that foundational self- supervised models can be effectively deployed via lightweight web frameworks to serve as reliable clinical decision- support systems.