A new AI framework grades acne severity by showing its work instead of just spitting out a number.
Researchers built CG-HAF, a system that combines two kinds of evidence instead of blending them into one opaque score. It averages severity predictions from several image classifiers trained on whole-face photos, then adds structured lesion data - count, detection confidence, and area - pulled from a separate object detector. A lightweight classifier fuses those two streams into a final grade on standard acne severity scales. On a widely used benchmark, the approach beat baselines that rely on whole-face evidence alone, with the biggest gains on the most severe cases.
Acne grading is notoriously subjective even among dermatologists, and black-box AI scores make that worse, not better, for anyone trying to trust or audit them. By keeping the whole-face and lesion-level evidence separate and legible, CG-HAF gives clinicians and skincare apps something closer to a paper trail than a verdict.
The catch: when the researchers tested it on a second dataset that uses a different grading standard, accuracy dropped, and the follow-up analysis pinned much of the blame on mismatched grading criteria rather than the model's eyesight - a reminder that interpretable doesn't automatically mean portable.