Good evening, and welcome to this February twenty twenty-six installment of our journal review, where we're working through the Journal of the American Academy of Dermatology. This issue actually gives us just one brief report to sink our teeth into today, and it's a substantial one for anyone thinking about the future of dermatopathology workflow, so we're going to spend real time with it rather than rushing past.

This is a brief report describing the development and validation of a deep learning framework for classifying erythematosquamous lesions on whole slide histopathologic images, out of Southern Medical University in Guangzhou.

Let's start with the clinical problem, because it's one you live with every time a biopsy report comes back hedged. Most erythematosquamous lesions get diagnosed clinically, but histopathology is the tiebreaker for the ambiguous cases, and you know exactly which ones those are: psoriasis versus eczema, early mycosis fungoides masquerading as spongiotic dermatitis, lichen planus-like interface changes overlapping with lupus. The authors frame the specific gap nicely. Current whole slide imaging diagnostics are bottlenecked by two things: they need an expert dermatopathologist to interpret them, and building a supervised model traditionally requires enormous amounts of hand-annotated data, meaning someone has to painstakingly label regions of interest across thousands of slides before the model can even start learning. That annotation burden is a major reason AI adoption in histopathology has lagged behind, say, dermoscopy.

So the methodological choice here is self-supervised learning, and it's worth pausing on why. In a self-supervised framework, the model learns representative features directly from the image patches themselves, without needing a human to pre-label what's important in each patch. The authors specifically built a hybrid convolutional neural network-transformer architecture, called CTransPath, which combines the local pattern-recognition strength of a convolutional network with the transformer's ability to weigh relationships across the whole slide. They compared this against two alternative architectures, one called HIPT, a hierarchical transformer built for gigapixel whole slide images, and a standard vision transformer, ViT, as a baseline. Choosing to benchmark against these two alternatives is a sensible design decision, since it lets them show the performance gain is coming from the hybrid architecture specifically, not just from throwing any deep learning model at the problem.

For the results: they trained and evaluated on close to thirteen hundred H&E stained whole slide images, spanning seven diagnostic categories, plaque psoriasis, eczematous dermatitis, pityriasis rubra pilaris, lichen planus, pityriasis rosea, lupus erythematosus, and patch or plaque stage mycosis fungoides, all confirmed by three board-certified dermatopathologists. The CTransPath model hit an overall accuracy of about eighty-nine percent, meaningfully outperforming both HIPT and the plain vision transformer, the latter of which lagged well behind at around sixty percent accuracy. Looking at the per-category numbers, performance was excellent for psoriasis and lupus, with precision essentially near-perfect for both, which makes sense given psoriasis dominated the dataset numerically. Performance was more modest for the rarer, more morphologically overlapping categories, pityriasis rubra pilaris and pityriasis rosea in particular, which is exactly what you'd expect given smaller sample sizes and genuine histologic overlap between those entities.

Importantly, they didn't stop at internal validation. They tested the model on an external cohort of just over two hundred slides from a geographically distinct hospital, Second Xiangya Hospital, covering psoriasis, eczematous dermatitis, and lichen planus. Accuracy held up at eighty-eight percent, essentially unchanged from the internal cohort, which is the kind of external validation you want to see before taking any of this seriously as a generalizable tool rather than a curve-fit to one institution's slide staining and scanning protocols. Sensitivity in that external set was strong for psoriasis at ninety-two percent, but dropped to the mid-seventies for eczematous dermatitis and lichen planus, so there's real room for improvement on the more subtle categories.

The finding I'd flag as most clinically interesting is the mycosis fungoides subset. We all know the diagnostic delay problem with early patch-stage MF being read as eczema, sometimes for years. Within their MF cohort, about half the cases were early-stage, and in the validation subset, the model correctly flagged three of four early MF cases as MF rather than eczema, while maintaining ninety-five percent specificity distinguishing full MF from eczema overall. That's a small numerator and denominator, four cases, so I want to be careful not to oversell it, but directionally it's an encouraging signal that a model like this could eventually assist with exactly the differential that generates the most real-world diagnostic anxiety and delay.

On limitations, the authors are appropriately modest, and there are some the paper doesn't spell out that are worth naming ourselves. This is not a randomized or even truly prospective clinical trial, it's a retrospective classification exercise against a reference standard set by three dermatopathologists at one institution, so any systematic bias in how those three experts called the ambiguous cases gets baked into the model's ground truth. The rarer categories, pityriasis rubra pilaris, pityriasis rosea, lichen planus, mycosis fungoides, all had quite small sample sizes, so those per-class metrics carry wide uncertainty even though the topline accuracy number looks robust. And the external validation, while reassuring, covered only three of the seven original categories, so we don't yet know how the model generalizes on lupus, mycosis fungoides, or pityriasis rubra pilaris outside the training institution.

So, practical takeaway. This is not practice-changing today, there's no product here you could deploy in your lab tomorrow, and the authors themselves frame it as a proof of concept aimed at assisting less experienced dermatologists and enabling triage in resource-limited settings, not as a replacement for expert dermatopathology read. But it is a genuinely interesting methodological direction worth tracking, specifically because the self-supervised approach sidesteps the annotation bottleneck that has slowed down AI adoption in our field, and because the early signal on distinguishing early mycosis fungoides from eczema touches one of the diagnostic problems that actually matters most to patient outcomes. I'd file this under promising early-stage AI research to watch for follow-up validation studies, particularly ones with larger mycosis fungoides cohorts and broader external testing, rather than anything to change your sign-out practice around.

That wraps our coverage for this episode. Thanks for spending the time with this one report in depth, and we'll be back with the next installment covering more from this issue.