A Skin Cancer Imaging Study Draws Its Second Round of Published Pushback in Three Months

A reply letter posted online October 14, 2026 in the Journal of the American Academy of Dermatology marks the second round of published disagreement over a single skin cancer imaging study in under three months. The reply, from Christoph Müller and Harald Kittler of the Medical University of Vienna, answers a comment from Fang Jiang, Yifan Xuan, Shiyuan Li, and Dan Zhao, posted the same day. Both pieces sit behind JAAD's paywall with no indexed abstract, so the specific objections raised by Jiang and colleagues aren't verifiable from the public record yet.

The paper under fire is Müller and colleagues' own March 2026 study, "Comparative and complementary diagnostic value of dermatoscopy and clinical close-up photography in skin cancer diagnosis," published in JAAD's July print issue (volume 95, issue 1, pages 70 to 77) and built on the MILK10k dataset. That study ran an online reader experiment: 283 participants diagnosed 1,567 paired clinical close-up and dermatoscopic images under the International Dermoscopy Society framework. The results, analyzed with a generalized linear mixed model, gave dermatoscopy 85.0 percent sensitivity against 74.2 percent for close-up photos alone, while specificity ran the other way: 66.8 percent for dermatoscopy against 71.9 percent for close-ups. Combining both modalities raised the odds of a correct diagnosis, by 52 percent for dermatoscopy and 40 percent for close-up photography.

The authors flagged their own limitation in the original paper: the reader study simulated teledermatology, meaning clinicians reviewed images on screen instead of examining patients in person. Selection and verification bias stay possible under that design. That caveat is the likeliest opening for outside critique, since remote image review and face-to-face exams don't produce identical diagnostic conditions and the literature comparing the two for melanoma detection specifically remains thin.

This is also not the first exchange the study has drawn. An earlier comment from Hafiza Syeda Bisha, F.N.U. Tooba, and Muhammad Hassan Saeed, of Jinnah Sindh Medical University and Jinnah Postgraduate Medical Center in Karachi, ran in JAAD on August 6, 2026 (DOI 10.1016/j.jaad.2026.08.002), with a same-day reply from Müller and Kittler (DOI 10.1016/j.jaad.2026.07.121). Two separate critiques landing within three months of each other points to a paper that provoked scrutiny on methodological grounds, even though the content of neither exchange is retrievable without journal access.

One layer sits underneath both rounds of correspondence: the MILK10k dataset itself. Tschandl and colleagues introduced it in the Journal of Investigative Dermatology in February 2026, describing 10,480 paired close-up and dermatoscopic images from 5,240 lesions across five centers, covering 48 diagnoses aligned to the ISIC-DX taxonomy, with 95.7 percent of cases confirmed by histopathology. The dataset is hosted openly on the ISIC Archive under a CC-BY-NC license, with a held-out 958-image benchmark subset reserved for testing AI models, where a hierarchical machine learning pipeline reached 53.6 percent top-1 accuracy and 96 percent specificity. Any critique of the reader study inherits whatever composition or labeling choices went into that underlying dataset, which the March 2026 paper didn't build from scratch.

What remains unresolved, for now, is which of those threads the October 14 comment actually pulls on: the teledermatology simulation, the statistical handling of paired reader data, or the dataset's own diagnostic mix. The comment and reply are both listed on JAAD's site00479-2/fulltext) and on PubMed for the original study, but full text sits behind institutional access, leaving the exact wording of Jiang, Xuan, Li, and Zhao's objection as the open question for anyone checking this correspondence firsthand.