A two-stage workflow for vitiligo diagnosis: clinical characteristic classification and large language model (LLM)-based report generation.

Background: Vitiligo is a common depigmenting disorder frequently misdiagnosed due to its visual similarity to other hypopigmentary conditions. While artificial intelligence (AI) has shown promise in dermatological image analysis, most models lack interpretability and fail to provide actionable clinical recommendations. Objective: To develop and validate an AI-assisted diagnostic system that integrates a large language model (LLM) for differentiating vitiligo from ten other hypopigmentary disorders, while providing interpretable characteristics and structured clinical reports. Methods: We retrospectively collected clinical images from a multicenter cohort, including patients diagnosed with vitiligo or one of ten other hypopigmentary disorders across five hospitals in China. A multi-task Vision Transformer model was trained to classify eight key clinical characteristics and output diagnostic probabilities. The model's structured predictions were then fed into the DeepSeek LLM to generate comprehensive clinical reports. Diagnostic performance was evaluated on an independent set of 175 images (87 vitiligo and 88 non-vitiligo) and compared with 43 dermatologists. Results: The study included 13,322 clinical images from 2,974 patients. For distinguishing vitiligo, the model achieved an AUC of 0.9906 (95% CI: 0.9844-0.9968), with sensitivity of 98.29% and specificity of 93.73%. Through multi-task learning, the model demonstrated accurate classification of eight key clinical characteristics, notably achieving 88.12% accuracy in identifying typical location and 86.78% in recognizing edge morphology. In a comparative test using 175 independent images, the AI model (AUC = 0.98) significantly outperformed a panel of dermatologists, particularly in diagnostic sensitivity. Moreover, the system successfully generated clinical reports via the DeepSeek LLM, providing structured diagnostic suggestions, differential diagnoses, and treatment plans. Conclusion: This multicenter study presents a clinically interpretable AI system for accurately discriminating vitiligo from similar disorders. By generating LLM-based reports, it enhances diagnostic transparency and supports clinical decision-making, offering potential assistance especially in resource-limited settings and advancing intelligent tools in dermatology.