Welcome back to the journal review — this is our April 2026 walkthrough of the Journal of the American Academy of Dermatology. Four pieces on the docket today, and it's a mixed bag: a systematic review and meta-analysis on cutaneous Kaposi sarcoma treatment, a research letter pushing back on GPT-4 Vision's dermoscopic performance, a letter to the editor on imaging in high-risk squamous cell carcinoma, and a back-and-forth correspondence on a dermoscopic sign for pigmented squamous cell carcinoma in situ. Let's get into it.

First up is a systematic review and meta-analysis looking at treatment efficacy for cutaneous, HIV-associated Kaposi sarcoma. The background here is straightforward but clinically real: Kaposi sarcoma remains the most common HIV-associated malignancy, and while antiretroviral therapy has cut incidence substantially, there's still no standardized treatment for limited, cutaneous-stage disease. Every one of us has faced that patient with a few stubborn violaceous plaques and no clear algorithm to follow.

Methodologically, the authors ran a PRISMA-guided search across PubMed, Embase, MEDLINE, and CENTRAL going back to 1946, which tells you immediately how sparse and old this literature is — you don't reach back eight decades unless the modern evidence base is thin. Eighteen studies met inclusion criteria, but only eleven had data clean enough to pool, and they used random-effects models with subgroup analysis by treatment modality — the right choice here, because you're combining wildly different agents, dosing schedules, and patient populations, and a fixed-effect model would falsely assume one true effect size across all of them.

On results: topical therapies — alitretinoin gel, halofuginone, docosanol — were the gentlest but weakest tools, with essentially no complete responses pooled and a no-response rate over seventy percent. Alitretinoin was the standout of that group but still only modest. Intralesional therapies were more of a mixed bag — intralesional TNF-alpha had strong lesion-level responses but meaningful systemic toxicity, interferon-alpha worked better in patients with preserved CD4 counts above 400 but was poorly tolerated in advanced HIV, and agents like bleomycin, vinblastine, and even human chorionic gonadotropin looked promising with low toxicity but simply haven't been validated in larger trials. Pooled complete response for intralesional approaches landed around 42%, with substantial heterogeneity.

Radiation therapy came out as the clear leader — fractionated regimens using higher total doses in the 20 to 40 Gray range consistently beat single-fraction approaches, while lower-dose regimens in the 8 to 20 Gray range were more for palliation than cure. Pooled complete response for radiation was around 63%, with the lowest no-response rate of any modality at roughly 7%, and this subgroup difference was statistically significant. Systemic HAART itself showed the single highest complete response rate in the pooled data, around 80%, but the authors are appropriately cautious flagging this, given how limited that dataset was.

The honest limitations here are substantial, and the authors say so plainly: heterogeneity was high across almost every outcome, risk of bias was low in about three-quarters of studies but moderate in the rest, and — critically — most included studies predate modern antiretroviral regimens, which the authors note now achieve response rates anywhere from 35% to 95% on their own, especially in combination. So this pooled analysis is largely describing a pre-modern-ART treatment landscape being layered onto a post-modern-ART disease epidemiology.

Practical takeaway: this isn't practice-changing in the sense of overturning what you already do, but it is a useful evidence consolidation. Radiation remains your most reliable local modality for a persistent cutaneous plaque, particularly using higher fractionated doses rather than palliative low-dose regimens, and optimizing ART remains foundational rather than adjunctive. The real message for us is that intralesional and topical options are reasonable for low-morbidity, low-burden lesions, but shouldn't be oversold — and that the field genuinely needs contemporary trials designed around patients who are already virologically suppressed, because that population barely exists in this pooled dataset.

Next, a research letter that's really an evolving-technology commentary rather than an original study — it's a direct response to a prior JAAD publication on GPT-4 Vision's dermoscopic accuracy. The original paper by Tadros and colleagues found GPT-4 Vision performed substantially worse than dermatologists at diagnosing skin conditions from dermoscopic images, with a particular tendency to overcall melanoma. This letter's authors make a sharp and important point: large language models change fast, and testing one snapshot version doesn't tell you much about the trajectory.

Their approach was a small but clever comparative exercise — melanoma versus benign, biopsy-validated dermoscopic images from the International Skin Imaging Collaboration dataset, one hundred images split evenly, tested across a sequence of GPT versions from the original GPT-4 Turbo through GPT-4o, o3, GPT-4.1, and GPT-5, all using chain-of-thought prompting. This is a reasonable design for what it is — a rapid benchmarking exercise, not a clinical validation study — and the authors are transparent that it's meant to illustrate a trend, not establish clinical readiness.

The findings: consistent with the original paper, earlier models overcalled melanoma badly, with specificity down in the 8 to 20% range and overall accuracy hovering around 50 to 54% — essentially coin-flip territory. But newer models improved accuracy by 15 to 20 percentage points, edging toward 70%, with much more balanced sensitivity and specificity. The trade-off worth flagging for a surgical audience: as specificity rose sharply, sensitivity fell by 20 to 35 percentage points, meaning newer models miss more melanomas even as they stop crying wolf on benign lesions — a trade-off you'd never accept in a screening tool. Adding few-shot learning — supplying five reference images per class alongside each query — pushed accuracy further, with bigger gains in the older, worse-performing models, but performance plateaued around 75 to 80% even in the best configurations, still short of published dermatologist-level performance.

The authors' bottom line, and I think it's the right one: don't judge a rapidly moving technology by a single frozen version, but also — and they say this explicitly — none of these models, old or new, are ready for clinical use. For us, this is interesting, not actionable. Nothing here changes triage or biopsy thresholds. File it as a technology-watch item, particularly the sensitivity-specificity trade-off, which is the exact wrong direction if you're trying to build a melanoma-detection aid rather than a melanoma-avoidance aid.

Third, a letter to the editor responding to a retrospective cohort study on radiologic imaging in high-risk cutaneous squamous cell carcinoma. The original study, published by Wei and colleagues, found that imaging could uncover previously unrecognized local invasion and nodal metastasis in high-risk cutaneous squamous cell carcinoma — findings that changed management in some patients. This letter is commentary, not new data, so there's no methods or results section to walk through — instead the correspondents raise several constructive critiques.

Their main points: the original study was single-institution and retrospective, so selection bias in who got imaged and why is a real concern, and they'd like to see multi-institutional validation before this becomes broadly generalized practice. They also flag that ultrasound was underexplored in the original study — they cite their own institutional experience and a Chinese cohort study in high-risk head and neck cutaneous squamous cell carcinoma showing ultrasound sensitivity around 73% and specificity around 94% for nodal metastasis, arguing that high-frequency ultrasound is a cheap, real-time, biopsy-guiding tool that deserves more attention alongside cross-sectional imaging. They also note the original paper only briefly touched on surveillance imaging after treatment, despite detecting subclinical recurrences on follow-up scans, and argue that's exactly where future prospective work should go — establishing actual timing and modality guidelines rather than ad hoc surveillance. Finally, they raise the more speculative point that combining imaging with molecular or genomic biomarkers could eventually refine risk stratification further, citing precedent from oncologic histology-genomic modeling.

None of this is practice-changing on its own — it's a thoughtful critique urging caution and further study rather than new evidence. But it reinforces something worth carrying into your own practice now: for your genuinely high-risk squamous cell carcinoma patients — deep invasion, perineural spread, immunosuppression — baseline imaging is increasingly defensible, and point-of-care or dedicated ultrasound is a reasonable, underused adjunct for regional nodal assessment, especially where cross-sectional imaging access is limited.

Last, a short correspondence exchange — really an author's response to a prior response — concerning the "plumage sign," a proposed dermoscopic clue for pigmented squamous cell carcinoma in situ. There's no methods or data here; this is a clinical pearl clarification. The original describer, Erin Hurd, is responding to a critique from a Memorial Sloan Kettering group who argued that what Hurd called the plumage sign might actually be indistinguishable from the "jelly sign" of solar lentigo.

Hurd's response is essentially a clarification of morphology: the plumage sign, as she defines it, consists of sharply demarcated, linear, pointed triangular arches with pigment concentrated at the apex — like feather tips — as opposed to the jelly sign's rippled, parallel, more rounded and semi-elliptical hyperpigmented borders seen in solar lentigines. She concedes the original photographs were suboptimal and didn't clearly annotate the relevant structures, which she believes drove the confusion. She adds anecdotally that she has since identified eight additional cases, all confirmed as pigmented squamous cell carcinoma in situ on pathology, and describes one case where recognizing the pattern prompted deeper pathology sections that upgraded a diagnosis from pigmented actinic keratosis to pigmented squamous cell carcinoma in situ.

This is purely descriptive correspondence — no systematic data, no controlled comparison, just an author sharpening a clinical observation and inviting collaborative validation with the MSKCC group. For practice, treat this as a pattern worth watching for on dermoscopy — sharp, acute-angled, feather-like pigmented arches with apical pigment concentration in a lesion you're already worried about — but it remains an anecdotal pearl, not a validated diagnostic criterion, and shouldn't yet change your biopsy threshold on its own.

That wraps our April 2026 review. A meta-analysis reinforcing radiation as your most reliable local option for cutaneous Kaposi sarcoma while underscoring how outdated the underlying trial data really are, a cautionary but evolving look at GPT models in dermoscopy that remain nowhere near clinical-grade, a thoughtful critique pushing for ultrasound and multi-institutional validation in high-risk squamous cell carcinoma imaging, and a dermoscopy pearl still working out its own definition in real time. Thanks for listening, and we'll see you next month.