iridis.

Perceptual skin-tone phenotyping from open dermatology imaging data — testing whether data-driven color clusters carry more structure than the clinical scales used to describe them.

Why this exists

Fitzpatrick skin type — six ordinal buckets, originally designed for burn-risk classification — is the de facto skin-tone standard in dermatology datasets, and by extension in the AI models trained on them. It's a coarse scale being asked to do a job it wasn't designed for. iridis approaches the question empirically instead of by argument: extract robust color features directly from images, let clustering discover the structure that's actually there, and compare how well that discovered structure predicts against how well the clinical labels predict, from identical features.

Approach

A layered masking pipeline isolates skin from background and pathology before any color is measured — class-agnostic foreground segmentation, then a ResNet18-U-Net (trained on ISIC2018 lesion masks) to exclude the lesion itself, with a center-crop fallback for the fraction of images where segmentation degenerates. Each image is downsampled, pixel-sampled, and converted to CIE Lab; per-image color is the median over sampled pixels rather than the mean, which is far less sensitive to residual artifacts and small segmentation errors. MiniBatchKMeans produces an initial fine-grained partition, and nearby clusters are merged using CIEDE2000 perceptual distance — so the final categories reflect what a human eye would distinguish, not an arbitrary cluster count.

Key results

On Fitzpatrick17k (~12.6K images) and ISIC 2018 Task 1 (~5.2K lesion-masked images), using identical Lab/LCh features, the discovered color clusters are roughly 2.5x more predictable than the clinical Fitzpatrick labels they're meant to summarize — 96% classification accuracy for the discovered clusters versus 36–42% for the six-bucket clinical scale, across both a classical (Random Forest) and an in-context (TabPFN) model. That gap is evidence the clinical scale is collapsing real, learnable perceptual variation, not a modeling artifact.

Status

Active. Core pipeline (masking, featurization, clustering, benchmarking) is implemented and reproducible. Segmentation is being reworked with additional data before an interactive demo is built.