YOLOv11-based segmentation of vitiligo lesions using standardized clinical photography

To develop and validate a YOLOv11-based deep learning framework that automatically segments vitiligo lesions in standardized clinical photographs captured under controlled conditions, enabling reproducible lesion quantification across longitudinal visits.

Standardized photographs from 18 patients were acquired at baseline, 3 months, and 6 months under controlled imaging conditions, including fixed lighting, camera settings, and color calibration. Images underwent preprocessing involving resizing, color and illumination normalization, noise reduction, and geometric standardization. Expert dermatologists generated binary lesion masks that were used as ground truth for training a one-stage YOLOv11 segmentation model. Model performance was evaluated using patient-wise data splits and 5-fold cross-validation.

The proposed model achieved 95 % accuracy, 92 % precision, 89 % recall, 90 % F1-score, with mAP@0.5 values of 0.878 (bounding box) and 0.892 (mask) from precision-recall analysis. Qualitative review indicated consistent segmentation quality across follow-up visits and moderate illumination variations, though formally stratified quantitative analysis by visit was not performed in this study. Visual assessment demonstrated close agreement between predicted and expert-annotated lesion boundaries, with minor under-segmentation observed mainly in low-contrast peripheral regions.

Standardized clinical photography combined with a YOLOv11-based segmentation framework enables reliable vitiligo lesion delineation and has potential for integration into quantitative disease monitoring and clinical decision-support systems.