Deep Learning-Based Segmentation of Lesions from Wide-Field Vitiligo Images

Abstract

Evaluation of vitiligo relies on accurate segmentation of lesions, and traditional segmentation methods mainly focus on near-field images. This study proposes a deep learning-based model for accurately segmenting lesions in wide-field vitiligo images. In this study, a dataset of 1267 wide-field vitiligo images was established to train and evaluate segmentation models. A Swin R-CNN model, which combined a Swin Transformer tiny network with a watershed algorithm, was proposed for segmenting lesions. The performances of the Swin R-CNN model and five other models were evaluated and compared through visual and quantitative perspectives. Additionally, the Spearman rank correlation test was performed to analyze result consistency between the Swin R-CNN model and dermatologists in measuring lesion area. The Swin R-CNN model accurately segmented lesions in the wide-field vitiligo images, surpassing other models in both visual and quantitative performance, with an average precision of 84.72% and an average recall of 77.81%. The correlation coefficients between the evaluation results of the Swin R-CNN model and three dermatologists were 0.88, 0.94, and 0.91, respectively. The Swin R-CNN model accurately segments lesions in the wide-field vitiligo images and quantifies lesion area at the dermatologist level. The Swin R-CNN model can provide reliable analytical results for the vitiligo evaluation.