Weakly-Supervised Vitiligo Segmentation

A weakly supervised framework for segmenting vitiligo lesions in skin images using saliency propagation on superpixel-based graphs.

Developed a novel weakly supervised segmentation framework for vitiligo lesion detection in skin images. The method leverages image-level labels to generate precise saliency maps via saliency propagation on superpixel-based graphs, avoiding the need for expensive pixel-level annotations.

Key features:

  • Robust against indistinct lesion borders and inconsistent illumination in user-captured photos taken with phone cameras
  • Curated Vit2019, a pixel-wise annotated vitiligo image dataset with more than 2,000 images, unprecedented in scale as of 2018

Published at IEEE BIBM 2019 with 25+ citations. This was my undergraduate research at Southeast University, supervised by Prof. Siyu Xia.

References