Identifying Cellular Heterogeneity in Vitiligo Using Deep Learning on In Vivo Reflectance Confocal Microscopy

Introduction Vitiligo is an autoimmune disease characterized by skin depigmentation, affecting approximately 1% of the global population. Reflectance confocal microscopy (RCM) enables high‐resolution visualization of pigmentary changes; however, accurate and efficient identification remains challenging due to blurred boundaries, variable brightness, and low contrast. Method We retrospectively analyzed RCM images from 105 vitiligo patients. The proposed pipeline incorporated brightness equalization, segmentation using a U‐KAN network with reverse learning, and multichannel feature fusion of grayscale, histogram of oriented gradients (HOGs), and L‐luminance channels to capture both morphological and luminance‐based features. Pigment ring subtypes were classified using a Data‐efficient image Transformer (DeiT) transformer model, and concept relevance propagation (CRP) was applied to interpret model attention. Results Feature enhancement through HOG and L‐channel fusion improved image contrast and detail visibility, yielding classification accuracy between 90% and 97%. The U‐KAN network achieved precise cell segmentation with an Intersection‐over‐Union (IoU) of 78 ± 2%. The DeiT model outperformed junior dermatologists and matched senior dermatologists in accuracy, precision, and specificity. Ablation studies showed the highest performance in the complete and incomplete categories, with recall reaching 100% for complete pigment rings. CRP maps indicated that contextual background information enhanced classification performance. Conclusion This study presents an innovative RCM processing pipeline integrating reverse learning and multichannel fusion, achieving dermatologist‐level performance in pigment ring classification. The approach demonstrates strong potential as an automated diagnostic tool for vitiligo, with future work aimed at improving generalization and clinical applicability.