Welcome back to the journal review. This is your walkthrough of the October twenty twenty-six issue of the Journal of the American Academy of Dermatology, and we've got four pieces worth your time today — an original deep learning study on facial pigmented lesions, a JAAD Game Changer retrospective on melanoma regression, a SEER-based brief report on who actually gets Mohs surgery for pleomorphic dermal sarcoma, and a sharp letter pushing back on how we interpret AI performance on clinical vignette challenges. Let's get into it.

First up is an original study: AI-assisted dermoscopy for lentigo maligna detection, out of Ankara University in collaboration with Imperial College London. The clinical problem here is one you live with constantly — distinguishing lentigo maligna from its benign mimickers, pigmented actinic keratosis and the solar lentigo/seborrheic keratosis spectrum, on chronically sun-damaged facial skin. The authors frame the underlying difficulty nicely: facial skin's flattened rete ridges produce a pseudonetwork pattern that degrades the specificity of our usual dermoscopic criteria, and even newer features like perifollicular linear projections, while fairly specific for early lentigo maligna, aren't very sensitive. They also nod to the "inverse approach" concept — that absence of a convincing benign pattern may be more diagnostically useful than presence of any single malignant feature. That's the conceptual gap an AI system is positioned to fill, since it isn't constrained to rule-based pattern recognition.

Methodologically, this is a retrospective study building and testing a convolutional neural network — specifically an Xception-based architecture — on 722 lesions and 894 dermoscopic images, split into lentigo maligna, pigmented actinic keratosis, and the combined solar lentigo/seborrheic keratosis group. They held out twenty percent of lesions as a completely independent test set, and trained on the remainder using five-fold stratified cross-validation. It's worth pausing on why that design choice matters. A single train-test split gives you one noisy estimate of performance that depends heavily on which cases happened to land in which bucket. Five-fold cross-validation instead rotates through five different train-validation splits of the same development data, averages the performance across all five, and gives you both a mean and a standard deviation — so you get a much more stable, less luck-dependent estimate of how the model actually generalizes. The "stratified" part means each fold preserves the same proportion of lentigo maligna to benign lesions as the overall dataset, which matters because these classes weren't perfectly balanced to begin with. Critically, they also kept lesion-level separation between folds and the held-out test set, so the same lesion's multiple images never leak across splits — a subtle but important safeguard against inflated performance.

For the human half of the study, twenty-six dermatology residents with an average of under three years' experience classified lesions first unaided, then again after seeing the AI's zero-to-one-hundred malignancy score. Performance before and after was compared using paired analysis — appropriate because the same residents, on the same cases, are being measured twice, so you want a test that accounts for that within-subject correlation rather than treating the two timepoints as independent samples. They also used Cohen's kappa alongside raw accuracy, which is worth teaching briefly: raw accuracy can be misleadingly high just by chance agreement, especially with imbalanced classes, so kappa corrects for the agreement you'd expect by random guessing alone and gives you a cleaner measure of true diagnostic concordance with the reference standard.

On results, the binary malignant-versus-benign model hit a mean test accuracy around eighty-four percent, with sensitivity around ninety percent and specificity around eighty-two percent. The three-class model, asked to separately call lentigo maligna, pigmented actinic keratosis, and the solar lentigo/seborrheic keratosis group, dropped to about seventy-one percent accuracy, which makes sense — finer-grained classification is a harder task. The real story, though, is the human-AI piece. Resident accuracy rose from roughly sixty-five percent at baseline to seventy-four percent with AI assistance, a highly significant improvement, and the benefit was concentrated exactly where you'd want it — lentigo maligna detection jumped about sixteen and a half percentage points, far more than the gains seen for the benign mimickers. Agreement with the reference standard, measured by kappa, moved from moderate to substantial. And when they broke down exactly how the AI changed resident answers, it flipped roughly three hundred wrong answers to right for every sixty-three it flipped from right to wrong — about a four-to-one favorable ratio, meaning the tool was net-helpful even accounting for the times it misled someone.

The authors are appropriately measured in discussion. This is retrospective, single-institution, and the reader study used residents only — not attendings or general dermatologists, so we don't know if experienced clinicians would show the same magnitude of benefit, since ceiling effects often blunt AI gains in expert hands. There's also no multimodal clinical context — no patient age, lesion duration, or history folded in, just the image. For your practice, I'd call this interesting and promising rather than practice-changing today. It's a nice proof of concept that AI-assisted scoring meaningfully boosts trainee accuracy specifically for the lesion type where missing the diagnosis has the highest stakes, which has real implications for residency training and triage support in busy pigmented lesion clinics. But with specificity still only in the low eighties, a meaningful share of benign lesions would still get flagged, and we have no data yet on how this performs against fellowship-trained Mohs surgeons or dermatologic oncologists, or how it would actually shift biopsy thresholds in your own hands.

Next is a JAAD Game Changer piece — and for those less familiar with the format, this is a short retrospective commentary, written by a different author than the original study, reflecting on how an already-published JAAD paper changed practice. The underlying paper here is Wagner and colleagues' twenty twenty-four single-institution cohort study on histopathologic regression in primary cutaneous melanoma. The controversy being addressed is a long-standing one — whether regression in a primary melanoma specimen is a reassuring finding or a worrisome one, since prior literature has gone both directions. The original study retrospectively reviewed nearly twelve hundred newly diagnosed melanomas requiring sentinel lymph node biopsy, diagnosed between twenty-ten and twenty-fifteen. Patients whose tumors showed regression had more favorable relapse-free survival, distant metastasis-free survival, and melanoma-specific survival, and regression was also associated with a higher likelihood of a negative sentinel node. Perhaps the most clinically interesting extension: even among patients who went on to develop unresectable disease, regression in the primary tumor was associated with better progression-free survival specifically on immune checkpoint inhibitor therapy — but not on targeted therapy or chemotherapy. The lesson the commentary distills is straightforward and clinically useful: regression should be read as a reassuring prognostic marker, both at the point of sentinel node biopsy eligibility and later, if a patient progresses and ends up on checkpoint inhibition. It's a nice reminder that a histologic finding we sometimes hedge on in pathology reports actually carries real prognostic and potentially predictive weight, particularly in the immunotherapy era.

Third, a brief report using the SEER database — Surveillance, Epidemiology, and End Results — looking at predictors of Mohs micrographic surgery utilization over wide local excision in localized pleomorphic dermal sarcoma, the entity formerly called malignant fibrous histiocytoma. The background gap is simple: we know Mohs achieves low recurrence and comparable survival to wide excision in this disease, but nobody had characterized who actually gets offered Mohs versus excision.

The design is a retrospective SEER-17 query, capturing patients with biopsy-confirmed localized pleomorphic dermal sarcoma treated between two thousand and twenty twenty-one, excluding anyone with regional or distant disease or an unknown surgical approach — four hundred sixty-seven patients total. This is a nice opportunity to walk through two statistical tools worth knowing well. First, they used standardized mean difference, rather than leaning only on p-values, to flag baseline imbalance between the Mohs and excision groups. The advantage of a standardized mean difference is that, unlike a p-value, it isn't sensitive to sample size — it's a pure effect-size measure of how different two groups are on a given variable, and a value above roughly zero-point-one is conventionally treated as a meaningful imbalance worth adjusting for. Second, for the actual predictor analysis, they ran multivariable logistic regression with backward elimination. Logistic regression is the right tool here because the outcome — Mohs versus wide excision — is binary; the model estimates the log-odds of receiving Mohs as a weighted combination of patient and tumor characteristics, and when you exponentiate those weights you get adjusted odds ratios, which tell you how much more or less likely a given group is to receive Mohs, holding the other variables constant. Backward elimination simply means they started with every variable that showed meaningful baseline imbalance and sequentially dropped the least significant one until only the independently meaningful predictors remained — they also checked variance inflation factors to confirm multicollinearity wasn't distorting the model.

The results: sixty-five percent of patients underwent Mohs, thirty-five percent wide excision. On multivariable analysis, two factors remained independently associated with receiving Mohs — annual income of eighty thousand dollars or more, with roughly a two-and-a-quarter-fold increased odds, and head-and-neck tumor location, with roughly a three-and-a-half-fold increased odds. Both were statistically significant and, frankly, both are clinically meaningful magnitudes, not just statistical noise.

The authors' interpretation is that this reflects real disparities — income and tumor location both independently shaping who gets tissue-sparing surgery, with lower-income patients likely facing barriers around awareness, geographic access to fellowship-trained Mohs surgeons, and perceived cost, while the head-and-neck preference for Mohs tracks sensibly with tissue conservation in cosmetically and functionally sensitive sites. Limitations are the ones you'd expect from any SEER brief report: retrospective design, income measured at the census-tract level rather than truly at the individual or household level, a modest sample size, and no data on tumor size, depth, or differentiation, plus no visibility into referral patterns or institutional Mohs availability, which surely play a role but aren't captured in the registry. This isn't practice-changing in the sense of telling you how to treat an individual patient — the oncologic equivalence of Mohs and wide excision for this tumor is already established elsewhere. What it does is document a measurable access disparity that's worth attention at the referral-pathway and outreach level, particularly if you practice somewhere serving a broad socioeconomic catchment.

Last is a letter — a comment responding to a prior JAAD publication by Dezoteux and colleagues on large language model performance on dermatology clinical challenges. This isn't a full original study, but the authors embedded a tight little experiment of their own to make their point, so it's worth walking through. Their concern is that large multimodal models evaluated on clinical-challenge-style questions might be succeeding largely by reading the text vignette, not by genuinely reasoning over the images — which obviously matters a great deal when we're trying to judge whether these tools have real dermatologic visual competence.

They pulled forty-four JAMA Dermatology Clinical Challenge cases, a hundred twenty-four questions total, and tested four models — GPT-o1, ChatGPT-5.2, Gemini 3 Pro, and Pixtral — under three conditions: text plus images, text alone, and images alone, running each condition three times per model. The finding is pretty stark. Accuracy with text-plus-images was essentially identical to text-only across every model — for ChatGPT-5.2, roughly eighty-eight percent either way, with no significant difference. Strip out the text and leave only images, though, and accuracy collapsed to somewhere in the forties to high fifties — still better than the twenty percent you'd expect from blind guessing on a five-option question, but a dramatic drop. Perhaps the most telling detail: when they swapped in completely irrelevant images instead of the real ones, accuracy barely moved, and most of the explanatory text that supposedly referenced the image was really just paraphrasing the written vignette rather than describing any actual visual finding.

There's no formal limitations section since this is a letter, but the generalizability caveats are obvious and the authors don't pretend otherwise — this is one journal's challenge format, a handful of models at a single point in time, and model behavior shifts fast. The teaching point they land on is one every reader of AI-in-dermatology literature should internalize: when a paper reports high large language model accuracy on an image-containing clinical vignette, that doesn't necessarily mean the model is doing real visual reasoning — it may just be very good at pattern-matching rich textual descriptions, and the image may be along for the ride. They propose that future benchmarks should separate "image-supplementary" vignettes, where the text alone basically gives away the diagnosis, from "image-essential" ones that truly require visual interpretation, which far better mirrors how you actually work up a patient in clinic. This isn't practice-changing for patient care, but it's an important critical-appraisal lesson for how skeptically to read the next AI diagnostic accuracy paper that crosses your desk — including, frankly, some of the facial lesion AI work we just covered a few minutes ago.

That wraps this month's review — a promising AI adjunct for lentigo maligna detection, a reminder that histologic regression is good news across the melanoma disease spectrum, a documented access disparity in surgical selection for pleomorphic dermal sarcoma, and a healthy dose of skepticism about what AI models are actually "seeing" when they ace a clinical challenge. Thanks for listening, and we'll see you next month.