AI for vitiligo: high accuracy on curated images, but not yet ready for clinic

Design

Systematic review and meta-analysis of 54 studies

Population

Imaging-based AI models for vitiligo diagnosis, classification and severity

Primary outcome

Diagnostic accuracy and lesion segmentation performance

Effect

Pooled AUC above 0.90; segmentation Dice score 0.84

This systematic review covered 54 studies of machine-learning and deep-learning models applied to images of vitiligo for diagnosis, severity scoring and lesion segmentation. Searches ran to October 2024.

The pooled diagnostic area under the curve exceeded 0.90 and lesion segmentation reached a Dice score of 0.84. Several studies reported that models performed as well as human experts. The authors note that larger, more diverse public datasets are needed, and that studies using full-body images are scarce.

Because most models were trained and tested on selected images, performance on darker or lighter skin types, different cameras and mixed skin conditions is unknown. Quantitative lesion tracking is the more useful near-term role, since objective extent can support treatment decisions and trials.

• Standardise clinical photographs: same light, distance and body site at each visit.
• Use a validated score such as VASI for extent if you track vitiligo.
• Treat AI output as support for, not a substitute for, clinical assessment.
• Ask any vendor which skin types and image sources the model was validated on.

Why it matters

Objective measurement of repigmentation is a long-standing gap, and imaging tools may help close it.

Don't overread it

Accuracy in curated image datasets does not show that these tools are validated for routine clinical use.

The statistics, in plain English

An AUC above 0.90 means a model separates vitiligo from non-vitiligo images well in the test sets. It does not tell you how it will do in a busy clinic with varied image quality, or on skin types missing from the training data.

Read the rest in the app

You have read your two free briefings this month. The app carries all 27 specialties, every morning, free — and this finding is waiting in it.

Get Daily Doctor — free

Scan to keep reading on your phone. No account needed to start.