Arioon
9/12/2025 · 2 min read · Arioon Team

Introducing Arioon — skin analysis that works on every skin tone

We're building a dermatology AI platform whose accuracy doesn't fall off a cliff for Fitzpatrick IV–VI. Here's how, and why we started.

Dermatology AI has a quiet, persistent problem: it doesn't work as well on darker skin.

For years, this has been treated as a "data" issue — train on more diverse faces and the gap closes. It's a comfortable framing because it makes the fix feel inevitable. Just collect more data, and the model will get there.

But the actual mechanics of the gap are not entirely a data problem. They're a measurement problem.

Why the gap exists

Most skin analysis pipelines work on raw colour channels — they ask "how red is this pixel?" or "how dark is this region?" — and convert those answers into health scores using thresholds calibrated on a population that was overwhelmingly light-skinned. A model trained that way will literally see darker skin as "darker than baseline" and report sun damage, pigmentation, or dark circles that aren't there.

Adding more melanin-rich training images helps the classifier learn what "normal" looks like for those tones, but it doesn't fix the underlying issue: the model has no first-class concept of whose normal a given face should be compared against.

What we do differently

Arioon makes that concept first-class. Every analysis starts by computing the Individual Typology Angle (ITA) — a dermatology-validated scalar derived from CIELAB L* and b* channels — at the nasal bridge. ITA maps cleanly to Fitzpatrick tone group, and we use that mapping to select a tone-group-specific baseline for every melanin-sensitive parameter.

The four parameters where this matters most — pigmentation, sun damage, dark circles, redness — get calibrated baselines per tone group. A Fitzpatrick V face is compared to FST V norms, not FST II norms.

What's next

Phase 1 is a complete CV-only pipeline that already corrects most of the gap. Phase 2 introduces ML-based detectors trained with adversarial ITA debiasing — a domain-classifier loss that explicitly penalises the model for being able to tell what tone group a face belongs to. That's a real architectural commitment to fairness, not a marketing one.

We'll be writing more on each of these in the coming weeks.