Segmenting Vitiligo on Clinical Face Images Using CNN Trained on Synthetic and Internet Images.

Area of Science:

• Dermatology
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate vitiligo diagnosis and severity assessment are vital for patient care.
• Current methods rely on subjective estimations of affected skin area, leading to variability.
• Facial vitiligo significantly impacts quality of life but has been overlooked in automated analysis.

Purpose of the Study:

• To develop an automated method for accurate facial vitiligo segmentation and area estimation.
• To address the challenge of limited clinical data for training deep learning models.
• To improve the precision of vitiligo severity assessment, particularly for facial involvement.

Main Methods:

• A convolutional neural network (CNN) modified from the fully convolutional network (FCN) architecture was utilized.
• A novel image synthesis algorithm was developed to generate realistic facial vitiligo images.
• The CNN was trained using a combination of internet-sourced images and synthetically generated images.

Main Results:

• Synthetic images significantly improved the performance of facial vitiligo lesion segmentation.
• The proposed algorithm achieved a low error rate of 1.06% in facial vitiligo area estimation.
• The automated method demonstrated superior accuracy compared to manual assessments by dermatologists and existing segmentation techniques.

Conclusions:

• The developed image synthesis algorithm effectively enhances CNN-based facial vitiligo segmentation.
• This AI-driven approach offers a more objective and accurate method for assessing vitiligo severity.
• The findings pave the way for improved diagnostic tools and treatment monitoring for patients with facial vitiligo.