Welcome to this January twenty twenty-six review of Dermatologic Surgery, the journal of the American Society for Dermatologic Surgery. We've got four pieces this month — an original study on artificial intelligence and Mohs frozen sections, a reconstructive case report out of Italy, a database-driven cohort study on a rare adnexal malignancy, and a practical commentary on an under-discussed intraoperative finding. Let's get into it.

First up is an original study, really framed by the authors as a "lessons learned" piece, looking at whether a deep learning model trained on permanent pathology can be applied to Mohs frozen sections to classify squamous cell carcinoma. The background gap here is straightforward: almost all of the artificial intelligence work in cutaneous oncology has been built and validated on formalin-fixed, paraffin-embedded — or FFPE — whole slide images, because that's where the digitized repositories exist. Frozen section repositories are scarce, expensive to build, and labor-intensive to scan and store. So rather than build a frozen-section-specific model from scratch, this group took a previously published FFPE-trained classifier and simply ran it on real Mohs frozen sections to see where it broke — with the explicit goal of using those failure patterns to guide future fine-tuning.

Methodologically, this is worth pausing on because it's a nice illustration of a "stress-test before you build" strategy. They trained their weakly supervised model on about seven hundred and fifty FFPE skin biopsy slides — a mix with and without squamous cell carcinoma, adjudicated by four dermatopathologists — generating whole-slide tumor probabilities along with tile-level attention heatmaps that show where the model is "looking." They then took that frozen model, without any retraining, and ran it in inference mode on fifteen Mohs-derived frozen section slides, all biopsy-proven squamous cell carcinoma cases, quality-checked by a Mohs surgeon and histotechnician as part of routine practice. The rationale for choosing a small, qualitative sample rather than a large validation cohort is essentially resource-driven — frozen section whole-slide scanning is expensive and slow — and the authors are explicit that this is meant to be exploratory, not a definitive validation study.

The results tell a clear story. On the FFPE test set, the model performed excellently, with an area under the curve of essentially point-nine-nine — about as good as these models get. Applied to Mohs frozen sections, that dropped to roughly point-eight, which is a real, clinically meaningful decline, not just statistical noise. More concerning was the pattern of that drop: specificity and positive predictive value were both perfect, but sensitivity and negative predictive value were poor. In plain terms, the model was very good at confidently calling something negative correctly, but it was missing a substantial fraction of true tumor-positive slides — exactly the wrong failure mode for a triage tool. When the authors dug into the attention heatmaps to understand why, they found the model handled slides with abundant tumor or purely normal tissue reasonably well. But in frozen sections with only scarce tumor foci — which, notably, made up over half their sample, reflecting later Mohs stages — the heatmaps went awry, either fixating on normal tissue in the false negatives, or, in slides it happened to classify correctly, highlighting the wrong structures entirely: inflammation, muscle, parotid gland, large nerves, larger vessels. In other words, when it got the right answer in these deeper specimens, it was frequently getting it right for the wrong reason.

The authors' interpretation is that these deep anatomic structures — muscle, nerve, gland, vessel — are essentially out-of-distribution data for a model trained only on relatively superficial biopsy specimens, where such structures are rarely represented. This is consistent with prior reports of poor FFPE-to-frozen generalization in other organ systems. The obvious limitation is sample size — fifteen frozen slides is not a validation cohort, and this is a qualitative, hypothesis-generating exercise, not a powered performance study.

Practically, this is not practice-changing — there's no product here to adopt or reject. But it is a useful cautionary signal for anyone watching the AI-assisted frozen section space: a model's excellent performance on permanent sections tells you very little about how it will behave on Mohs stages that go deep, and the failure mode to worry about is exactly the one you'd fear — missed tumor in scarce, deep-tissue frozen sections, the very stages where accurate real-time interpretation matters most.

Next, a case report with a technique twist, out of a plastic and reconstructive surgery group in Rome — "Castañares two-point-oh: Recycling a Direct Brow Lift Into a Transverse Supratrochlear Artery Propeller Flap for One-Stage Nasal Reconstruction." The setup is a genuinely elegant clinical problem: an eighty-three-year-old man with a history of parotid basal cell carcinoma excision complicated by injury to the frontal branch of the facial nerve, left with brow ptosis and ipsilateral lower lid ectropion, who then developed a new suspicious lesion on the ipsilateral nasal sidewall and glabella. So the surgical plan had to solve three problems in one sitting: excise the new skin cancer, correct the brow ptosis, and correct the ectropion.

The clever part is how they linked the first two. Rather than performing a standard Castañares direct brow lift and discarding the excised redundant tissue — as you normally would — they islanded that tissue on the ipsilateral supratrochlear pedicle, using Doppler mapping to localize the vessel, preserved a cuff of corrugator muscle for collateral flow, fully skeletonized the pedicle through a fenestration in the orbital periosteum for safe rotation, and then rotated the resulting flap ninety degrees as a propeller to resurface the two-and-a-half by two-and-a-half centimeter nasal sidewall defect — with the donor site closing primarily, concealed along the brow hairline. Essentially, tissue that would otherwise be thrown away became a single-stage substitute for what would traditionally require a two- or three-stage paramedian forehead flap. Ectropion was addressed separately with a tarsal strip canthoplasty. At six months, the patient had a well-integrated flap, good aesthetic outcome, and functional correction of both the brow ptosis and the ectropion, with no complications and clear margins on final pathology.

The authors are upfront about the limitations of this approach as a generalizable technique, not just as a case. Harvesting a skin paddle above the brow risks eyebrow deformation or elevation, so if the donor site needs primary closure, this really only works for small defects — they specify up to about two-and-a-half centimeters — in the nasal root, sidewall, glabella, or periorbital region, and ideally in patients who already have some brow ptosis to "spend," whether from facial palsy or, in non-paralyzed patients, by performing a matching contralateral brow lift for symmetry. They also note a risk of mild bulkiness at the pivot point if the pedicle isn't fully skeletonized, which may need a minor secondary revision.

The takeaway here is squarely in the "interesting technique pearl" category rather than practice-changing — it's a single case. But conceptually it's worth banking: when a patient has a facial palsy-related brow ptosis or similar and an adjacent nasal defect, this Gillies-inspired "don't discard living tissue" approach can convert two separate procedures into one, single-stage reconstruction, sparing an elderly, comorbid patient additional surgery.

Now to the piece with the most direct bearing on surgical decision-making: a retrospective cohort study on factors associated with metastasis in eccrine porocarcinoma, using two national databases. The background gap is real and persistent — porocarcinoma is rare, existing surgical margin data is inconsistent, and there's no established evidence base for when to pursue sentinel lymph node biopsy, even though nodal metastasis clearly worsens survival.

For methods, the authors queried the National Cancer Database for cases diagnosed between two thousand four and twenty nineteen, and separately queried SEER — the Surveillance, Epidemiology, and End Results database — for cases from two thousand to twenty twenty. The reason for using two databases rather than one is methodological necessity: disease-specific survival isn't calculable in the National Cancer Database, so SEER was brought in specifically for that outcome. For the metastasis analysis, they built two multivariable logistic regression models — one excluding lymphovascular invasion so they could use the full cohort without losing patients to missing data, and a second, smaller model that added lymphovascular invasion in the subset where it was actually recorded. That's a sensible way to have it both ways: preserve statistical power in the primary model while still formally testing the variable most people would guess matters. For the sentinel node question, since patients who received a sentinel biopsy weren't randomized to get one, they used propensity score matching to build comparable groups before comparing survival — an appropriate, if imperfect, attempt to control confounding in an observational comparison.

The results are genuinely clinically useful. Out of thirteen hundred fifty-one patients, regional nodal disease was found in about six percent, and distant metastasis in about one percent. Tumor size tracked closely with metastatic risk — mean size was roughly double in node-positive patients compared with node-negative, around forty millimeters versus twenty. In the regression model, each additional millimeter of tumor size increased the odds of metastasis by about three percent, a statistically significant and clinically sensible dose-response relationship. Lymphovascular invasion was the single biggest signal in the whole paper — present in only a subset of tumors, but when present, forty percent of those patients already had metastatic disease at diagnosis, compared with under six percent when it was absent — roughly a seven-fold increase in the odds of metastasis, both statistically and clinically striking. Looking specifically at cumulative sentinel node positivity by tumor size, the rate crossed the five percent threshold above seventeen millimeters, and among all tumors larger than that cutoff, the overall risk of a positive sentinel node was about seven percent. Survival numbers put this in perspective: five-year overall survival was seventy-one percent with localized disease, forty-five percent with regional nodal involvement, and zero percent with distant metastatic disease — a stark, clinically obvious cliff that underscores why catching regional disease early matters. Sentinel node biopsy itself was associated with better overall survival in the primary Cox model, but after propensity matching to control for confounding, that benefit was only borderline significant — worth knowing, but not a slam dunk. In the SEER cohort, disease-specific mortality was numerically lower with sentinel biopsy but didn't reach significance, likely underpowered. The most dramatic finding in the whole discussion section was in the subgroup with confirmed regional nodal disease: those who went on to complete lymph node dissection had a disease-specific mortality of about sixteen percent at five years, compared with sixty-five percent in those who didn't — a large, statistically significant difference.

The authors are appropriately cautious about the limitations inherent to registry-based work: substantial missing data, particularly for lymphovascular invasion status, no centralized pathology review, and likely selection bias in who gets offered a sentinel biopsy in the first place — sicker or older patients may simply be less likely to be selected for nodal staging, which could inflate the apparent survival benefit even after matching.

For practical takeaways, this one earns a "practice-informing, verging on practice-changing" label for the size and lymphovascular invasion thresholds specifically. A tumor over roughly seventeen millimeters, or any tumor with lymphovascular invasion on pathology, should prompt a real conversation about sentinel lymph node biopsy. And if nodal disease is found, the completion lymphadenectomy survival signal is compelling enough to support referral for definitive nodal surgery. The claim that sentinel biopsy itself independently improves survival is softer — interesting and directionally supportive, but still short of definitive given the borderline statistics after matching.

Last is a shorter, non-data-driven commentary — "Peri-Tumoral Varicosities: Surgical Implications" — which reads as a clinical pearl piece rather than a study, so there's no methods or results section to walk through; it's built entirely around raising awareness of a finding and offering practical operative guidance. The premise is that bulky cutaneous tumors can behave like any other cause of venous outflow obstruction, generating peritumoral varicosities through some combination of four proposed, non-mutually-exclusive mechanisms: direct intraluminal venous invasion by tumor, a tumor-induced procoagulant local environment promoting thrombosis, straightforward mechanical compression of the venous system, or a high-flow state from increased tumor vascularity. The authors are candid that the literature has never actually parsed out which mechanism predominates in which tumor types or locations, and they flag this as an open area needing study — including whether the finding carries independent prognostic weight, given that vascular invasion in general is associated with survival rates below fifty percent in several malignancies.

Where this piece is genuinely useful is the operative playbook it offers. Preoperatively, local anesthetic with epinephrine around and beneath the tumor to induce vasoconstriction and cutaneous pallor is the first move, though the authors note honestly that achieving that pallor can be surprisingly difficult in these cases because of altered local fluid dynamics — tumescent anesthesia is offered as an alternative worth considering. Intraoperatively, they favor suture ligation over electrocautery for any vessel over about two millimeters, specifically to avoid the thermal injury, delayed healing, and worse cosmetic outcome that comes with cautery on larger vessels — clamping for temporary control and assistant-applied direct pressure to localize an obscured bleeding source are both recommended as adjuncts. For particularly vascular tumors, they describe a segmental extirpation approach — achieving hemostasis at set points during excision rather than all at once, which in the Mohs context might mean starting with the most varicose peripheral aspect of the layer first, at the cost of added operative time. Because these defects and their bleeding risk make it unwise to leave a Mohs wound open during tissue processing, they recommend the modified winch stitch — a pulley-based running suture and hemostat technique — for temporary approximation, followed by pressure-generating closure with horizontal mattress or running locked sutures. Postoperatively, the advice is a forty-eight hour pressure dressing and counseling patients to avoid Valsalva-type activities that raise venous pressure. They close by noting that while post-surgical bleeding complications overall run around one percent, when they do occur they're disproportionately linked to necrosis, infection, and dehiscence — which is the whole rationale for taking this finding seriously up front.

The practical takeaway here is low-cost and immediate: recognizing peritumoral varicosities preoperatively should trigger a mental shift toward anticipating a harder hemostasis case — favoring suture ligation over cautery for larger vessels, considering a segmental excision strategy, having the winch stitch technique ready for temporary closure, and setting expectations with the patient about pressure dressings and activity restriction afterward. It's not a data-backed, outcomes-proven protocol, but it's a sensible, easily adopted awareness point for the next bulky, boggy tumor you encounter.

That wraps our four articles for January. In short: an honest look at where an AI classifier trained on permanent sections falls apart on real Mohs frozen tissue, a clever single-stage reconstructive trick worth filing away for the right brow-ptosis-plus-nasal-defect patient, solid registry evidence that tumor size over roughly seventeen millimeters and lymphovascular invasion should be driving your sentinel node conversations in porocarcinoma, and a practical reminder to read peritumoral varicosities as an early warning sign for a tougher hemostasis case ahead. Thanks for listening, and we'll see you next month.