What a New Skin-Cancer AI Paper's Author List Reveals

On August 10, a short communication appeared online in the Journal of Investigative Dermatology under a title that reads like a fix: generative skin tone augmentation reduces Fitzpatrick skin type disparities in skin tumor classification. The paper, indexed on PubMed the same day, describes using image to image translation, a technique that resynthesizes an existing photo with different skin tone characteristics, to widen the range of skin tones a tumor classifier is trained on. That much is documented in the journal's own record: a peer reviewed short communication, revised in mid-July and accepted July 21, with keywords including data augmentation, deep learning, and skin of color.

What the wire summary leaves out is who wrote it. Public affiliation records list nine authors. Seven work in dermatology, biomedical engineering, or convergence medicine at Asan Medical Center and the University of Ulsan College of Medicine in Seoul, a hospital whose own patient population is overwhelmingly East Asian, a demographic that clusters in the lighter to medium Fitzpatrick range rather than the darker types the paper's title addresses. The other two authors sit outside Korea entirely: one is based at Our Lady of Fatima University in Valenzuela City, in the Philippines, and the other, Messay Tesfaye, is a dermatologist at Addis Ababa University's College of Health Sciences in Ethiopia.

Why the roster matters

The problem this paper is aimed at is not new. In 2018, dermatologist Adewole Adamson and software engineer Avery Smith argued in JAMA Dermatology that the most heavily cited training sets behind early skin cancer algorithms held too few images of darker skin to support confident diagnosis across all patients. That critique reshaped how the field talks about dermatology AI, and the fix usually proposed is more real images of skin of color, collected at the source. A generative approach tries something narrower and faster: synthesizing plausible variation from what a hospital already has on file, rather than waiting on new collection.

Co-authorship with a dermatologist practicing on the population a dataset is missing is one way to keep that shortcut honest, since it puts someone with direct clinical exposure to the target skin tones in a position to judge whether the synthetic images look right. Whether that review happened here, and how rigorously, is not something the public record currently shows.

What the record doesn't yet show

As of this writing, the article carries no public abstract on PubMed or its publisher's site, and the full text sits behind a subscription. That means the accuracy figures, the generative model's architecture, and the size of the tumor classification dataset are not yet independently checkable. A short communication is also, by design, a compact report rather than a full validation study, so any claimed reduction in disparity is more plausibly a first look than a settled result.

For readers following AI in dermatology, two questions are worth carrying into the fuller version of this study, whenever its data becomes public:

• Does the classifier's accuracy improve on real photographs of darker skin, not only on the synthetic images built to resemble them?
• Did a dermatologist who treats that population confirm the synthetic images were clinically plausible before the model was trained on them?

Sources: PubMed · JAMA Dermatology