Researchers Kaiqiao He, Tianle Xu, Yining Feng and others developed a two-stage diagnostic system to differentiate vitiligo from ten other hypopigmentary disorders. The study, published in Frontiers in Immunology on June 1, 2026, addresses common diagnostic errors by integrating a multi-task Vision Transformer (a neural network architecture processing images as sequential patches) with the DeepSeek LLM. The research utilized a dataset of 13,322 clinical images collected from 2,974 patients across five hospitals in China.

The system first classifies eight key clinical characteristics before feeding structured predictions into the DeepSeek LLM to generate comprehensive clinical reports. The model achieved an AUC (a metric measuring classification performance) of 0.9906, with a sensitivity of 98.29% and a specificity of 93.73%. Specifically, it reached 88.12% accuracy in identifying typical location and 86.78% in recognizing edge morphology.

In comparative testing against 43 dermatologists using 175 independent images, the AI model achieved an AUC of 0.98, outperforming human clinicians particularly in diagnostic sensitivity. The system provides structured diagnostic suggestions, differential diagnoses, and treatment plans, offering a transparent, interpretable framework intended to assist clinicians, particularly in resource-limited settings.

Researchers Kaiqiao He, Tianle Xu, Yining Feng and others developed a two-stage diagnostic system to differentiate vitiligo from ten other hypopigmentary disorders. The study, published in Frontiers in Immunology on June 1, 2026, addresses common diagnostic errors by integrating a multi-task Vision Transformer (a neural network architecture processing images as sequential patches) with the DeepSeek LLM. The research utilized a dataset of 13,322 clinical images collected from 2,974 patients across five hospitals in China.

The system first classifies eight key clinical characteristics before feeding structured predictions into the DeepSeek LLM to generate comprehensive clinical reports. The model achieved an AUC (a metric measuring classification performance) of 0.9906, with a sensitivity of 98.29% and a specificity of 93.73%. Specifically, it reached 88.12% accuracy in identifying typical location and 86.78% in recognizing edge morphology.

In comparative testing against 43 dermatologists using 175 independent images, the AI model achieved an AUC of 0.98, outperforming human clinicians particularly in diagnostic sensitivity. The system provides structured diagnostic suggestions, differential diagnoses, and treatment plans, offering a transparent, interpretable framework intended to assist clinicians, particularly in resource-limited settings.