Dermatology Data Collection for AI: What It Takes to Get It Right
Over the years I’ve worked with medical datasets across multiple specialties, and dermatology remains one of the most demanding areas for AI development. On the surface it can seem straightforward: collect images, label conditions, train a model. In practice, every stage requires careful planning and specialized expertise. The gap between a dataset that looks good on paper and one that actually performs in clinical settings comes down to how rigorously these challenges are addressed from day one.
There’s a version of this conversation that happens at conferences and in product pitches, and then there’s the version that people in the field actually live. This post is the second one.
If you’re building a dermatology AI model – whether you’re a data engineer sourcing training sets, a startup validating your pipeline, or a clinician trying to understand how this data ends up in software – the honest answer is that dermatology is one of the harder domains in medical AI. Not the hardest. But harder than it looks, and harder than most teams account for.
Here’s what the work actually involves.
How the Data Gets Captured
Dermatology data comes in from two main directions, and they’re not interchangeable.
Clinical photography is the most common starting point – a clinician using a phone or standard camera to document a lesion, rash, or anomaly. It’s fast, it generates volume, and many clinics already have informal workflows built around it. The images are usable, but they’re also inconsistent: different lighting, angles, distances, and device quality across every capture.
Dermatoscopy is a different category entirely. A dermatoscope is a handheld optical instrument – essentially a high-magnification skin microscope – that illuminates the surface layers of skin and reveals patterns, pigmentation, and structures that are completely invisible to the naked eye. The difference from a phone photo isn’t just quality; it’s clinical information. For conditions involving pigmentation or subsurface morphology, dermatoscopy isn’t optional.
Both types are worth collecting, and both should be labeled with their capture method so models can be trained and evaluated on each appropriately. Clinical photos are easier to source at scale. Dermatoscopic images are harder to acquire but carry significantly more diagnostic signals.
For higher-stakes cases – particularly where oncology is a concern – data collection goes further. Digital pathology labs can contribute whole slide images and tissue specimens. Oncology-specialist clinics can provide slides for suspicious cases where malignancy needs proper characterization. This isn’t a routine collection workflow; it requires dedicated specialist partnerships. But for conditions like melanoma, it’s the difference between a superficial dataset and one that actually reflects clinical reality.
Common Conditions vs. Rare Ones: The Volume Problem
Some dermatological conditions are straightforward to source at scale. Acne is among the most prevalent skin conditions globally – you can build a well-documented, accurately labeled dataset without much friction. The challenge there is quality control and demographic diversity, not volume.
Others are a different matter. Alopecia, particularly in its less common forms, is rare enough that building a statistically meaningful dataset requires active cooperation with specialist clinics over an extended period. General hospital networks won’t get you there.
The conditions where structured data collection can realistically support AI development include: contact dermatitis, atopic dermatitis / eczema, seborrheic dermatitis, melanoma, alopecia, rosacea, acne, psoriasis, and many more.
That list spans both high-prevalence conditions like acne and psoriasis, and genuinely rare or complex ones like alopecia and melanoma. Getting the rarer conditions right demands specialist partnerships – not just general dermatology departments, but oncology units and digital pathology labs equipped to handle suspicious cases. Without those relationships, the rare-condition tail of your dataset will be too thin to train on reliably.
Anonymization: More Layers Than You’d Expect
Medical data anonymization is already a rigorous process. In dermatology, there’s an additional layer of complexity that doesn’t get enough attention: a lot of dermatological data involves the face.
Facial acne is the obvious example. To train a model on facial skin conditions, you need images of faces – and images of faces are among the most sensitive categories of biometric data. Even a heavily acne-affected face that might seem visually distinctive enough to obscure identity still contains identifying features: the shape of the nose, jawline structure, the ears, distinctive bone structure.
Proper anonymization means removing or obscuring every part of the face that isn’t the affected area itself. If a patient has acne covering most of their forehead and cheeks, the regions of interest may be large, but the surrounding identifiers still need to go. Eyes, ears, and facial landmarks get removed, blurred, or replaced – in my experience, this often requires image-by-image judgment calls that no automated tool handles consistently.
That last point is important: you can’t fully automate this. After automated processing, human anonymization specialists need to review each image and confirm that no identifying information remains. The standard I hold to is simple: a qualified person looks at the image and signs off. Anything short of that is a liability.
This matters beyond ethics. How well anonymization was done directly determines the dataset’s legal usability across jurisdictions. GDPR and HIPAA have real enforcement teeth, and a dataset that passed automated checks but wouldn’t survive a proper legal review is not actually a usable dataset.
The Skin Tone Diversity Problem
Dermatology AI trained predominantly on light-skin images fails on darker skin tones – not marginally, but substantially. A landmark study published in Science Advances found that state-of-the-art dermatology AI models, when tested against a dataset with diverse skin tones, saw ROC-AUC scores drop by 27 to 36 percent compared to their original benchmarks. All models performed worse on dark skin tones and on uncommon diseases. Critically, even the dermatologists who labeled these datasets performed worse on dark skin images – which means the bias is embedded at the annotation layer, not just the image layer.
In a clinical context, that translates directly into missed or delayed diagnoses for patients with darker skin. This isn’t a statistical curiosity; it’s a patient safety issue.
The fix is straightforward in concept but demanding in execution: training data needs to reflect the actual skin tone distribution of the populations the model will serve. That means deliberately sourcing data across ethnic backgrounds and geographic regions – not just working with whatever cases happen to appear in one hospital network.
ISIC (the International Skin Imaging Collaboration), the organization behind the most widely used public dermatoscopy benchmarks in the field, has made dermoscopic image standardization central to its mission precisely because these variables matter so much. Their archive has grown to over a million images. But the diversity problem in publicly available data remains real, and for models intended for broad clinical use, teams need to go further than what’s publicly available.
Regional diversity matters beyond skin tone as well. Dermatological conditions can present differently depending on climate, sun exposure, environmental factors, and local disease prevalence. A model trained entirely on data from Northern European clinical sites will behave differently – sometimes very differently – when deployed in South Asia or Sub-Saharan Africa.
What Good Annotation Actually Requires
Annotators need to reliably distinguish between conditions that can look visually similar on camera but require different clinical responses. Seborrheic dermatitis and psoriasis can overlap in presentation. Tinea and contact dermatitis can be confused. When labeling errors are introduced by under-qualified annotators, they propagate into training data, and training data errors propagate into model behavior – at scale, across every inference the model makes.
The minimum standard here isn’t a medical degree. It’s clinical experience in dermatology specifically, and enough of it that the annotator has seen the full range of presentations each condition can take. Five or more years of specialist practice is a reasonable threshold. For oncology-adjacent cases involving melanoma or other skin cancers, even that bar may need to be higher.
Conclusions
Properly executed dermatology data collection is not a background task. It’s a genuine operational undertaking that spans clinical partnerships, data privacy compliance, image quality control, expert annotation, and demographic diversity sourcing – all simultaneously, and at scale. For a quick overview of how to evaluate whether a healthcare dataset is ready for AI development, I put together this short: Three Ways to Tell If Your Healthcare Data Is AI-Ready.
Teams that underestimate any one of those dimensions tend to end up with datasets that look complete on paper and underperform in the field. The most common failure modes are consistent: a rare-condition tail that’s too thin to train on, anonymization that passed automated checks but wouldn’t survive proper legal review, and demographic skew that makes the model accurate on populations represented in training and unreliable on everyone else.
None of these problems are unsolvable. But they require treating the data work with the same rigor as the model architecture work.
