Dermatology
Dermatology is an important test case for clinical AI because the same image can support a narrow referral decision, a broad differential diagnosis, patient education, or longitudinal monitoring. Those are different tasks with different reference standards and harms. Evidence from curated image studies should therefore not be presented as evidence that an algorithm improves patient outcomes or can replace clinical assessment.
After reading this chapter, readers will be able to:
- Distinguish lesion-triage evidence from evidence of diagnostic or patient benefit
- Interpret regulatory status and intended use for dermatology devices
- Evaluate image-based AI across skin tone, image capture, disease subtype, and care setting
- Counsel patients about consumer dermatology tools without equating information support with safe self-triage
- Assess emerging AI approaches for atopic dermatitis, psoriasis, and acne severity monitoring
Introduction
In routine care, a lesion classifier, a clinician-facing image system, an asynchronous teledermatology service, and a patient-facing informational tool have different intended users and failure modes. Evaluation should match the evidence to the decision being supported, then require prospective evidence before claiming that a system improves care.
Suspicious Lesion Assessment
Skin cancer AI should be assessed as a clinical decision-support problem, not as a contest against clinicians on a curated image set. The relevant questions are whether a tool improves referral or biopsy decisions in its intended population, what additional false-positive burden it creates, and whether performance generalizes across image capture, anatomic sites, disease subtypes, and patient groups.
Evidence Beyond Benchmark Claims
The early literature established that image classifiers can approach clinician performance on constrained lesion-classification tasks. More recent realistic reader studies are more informative for implementation. In a multi-institutional study of 1,117 clinical and dermoscopic cases, a foundation model exceeded readers with less than three years of dermatology experience, while dermatologists with more than 10 years of experience achieved the highest multiclass accuracy (Anriot et al., 2026). Image-reader performance is not prospective deployment or evidence of better patient outcomes.
A 2025 systematic review and meta-analysis of studies comparing AI with family physicians and dermatologists reported pooled melanoma sensitivity of 0.86 and specificity of 0.94. However, 25 of 38 included studies had high risk of bias, most often because the selected cases did not represent outpatient populations (Nadour et al., 2025). Spectrum bias matters: a system evaluated on selected lesions cannot be assumed to perform similarly in routine primary care, teledermatology, or patient-submitted photographs.
A 2026 systematic review restricted to prospective dermoscopy studies included 11 studies with more than 2,500 patients. Pooled sensitivity and specificity were 78.6% and 75.2% for dermatologists and 80.9% and 75.6% for AI alone, with no significant pooled difference. Only one included study evaluated dermatologists assisted by AI, and most studies had high risk of bias because lesions were preselected and classification tasks were narrow (Laiouar-Pedari et al., 2026). The prospective evidence supports comparable diagnostic performance in selected dermoscopy settings, not autonomous use or proven patient benefit.
Equitable Image-Based AI
Skin tone, disease morphology, image capture, and care setting are inseparable sources of uncertainty in dermatology AI. A claim that an algorithm was tested “across Fitzpatrick types” is insufficient by itself. Fitzpatrick skin type was developed to describe response to ultraviolet exposure, not to measure skin tone, race, ethnicity, or the optical properties of a clinical image.
In a prospective comparative study of skin-tone labeling methods, no assessed approach generalized reliably across in-person assessment, clinical photographs, and dermoscopy images. The study concluded that Fitzpatrick skin type is not a proxy for skin tone (Weir et al., 2025). This does not supply one universal replacement scale. It does establish that image-based dermatology studies should justify their measurement method and report how labels were obtained.
Minimum Evidence for Equity Claims
An adoption dossier should report:
- The intended population, clinical task, image modality, acquisition protocol, and reference standard
- Disease subtype and anatomic-site representation, including acral and amelanotic lesions when relevant
- A prespecified and justified skin-tone measurement approach, rather than a race label or Fitzpatrick type alone
- Group-specific sensitivity, specificity, calibration, and uncertainty, with sample sizes and confidence intervals
- External validation across sites and capture conditions, followed by post-deployment performance monitoring
The implication is practical rather than rhetorical: a vendor that cannot characterize its data and subgroup performance has not supplied enough evidence for equitable deployment. Broader population-health implications of representation and measurement bias are addressed in the Public Health AI Handbook’s ethics and equity guidance.
Synthetic Images Are Not a Remedy
Generative models can create dermatology-like images, but synthetic imagery should not be treated as a clinical illustration, training datum, or evidence of diversity without expert review and provenance. An experimental assessment of four image generators found that three of four generators underrepresented dark skin (3.9–8.7% of images; Adobe Firefly matched US demographics at 38.1% but had the lowest diagnostic accuracy, 0.94%), and only 15% of images were identifiable as the prompted condition (Joerg et al., 2025). Synthetic augmentation requires the same scrutiny as any other data intervention: it can reproduce or amplify the errors and gaps of its source data.
Patient-Facing Tools and Teledermatology
Patient-facing dermatology AI has three distinct uses: educational information, asynchronous image submission to a clinician, and automated diagnosis or triage. They should not be evaluated or communicated as interchangeable.
Informational Tools Are Not Self-Triage Tools
A Google-authored randomized survey of 2,345 US adults reading retrospective cases found that a prototype AI interface raised condition-naming accuracy from 7.9% to 22.8% but did not improve next-step accuracy (Sayres et al., 2026). This supports a narrow role for condition information and question preparation. It does not validate self-diagnosis, cancer exclusion, or treatment recommendations.
Two online reader experiments (623 lay people; 153 primary-care physicians) found that fairness-constrained dermatology AI improved accuracy, but multimodal LLM explanations raised lay automation bias: accuracy rose when the model was right and fell when it was wrong (Xu et al., 2026). PCPs were more resistant to incorrect advice, but putting AI first with LLM text produced the largest rise in full deference (+19.0%, in small PCP arms of 17 and 26 users). The studies used 12 images without clinical history, the first author holds a part-time visiting faculty role at Google, and the authors favor a Human-First workflow rather than consumer self-diagnosis.
Patients should be advised that an app result must not rule out malignancy or defer clinically indicated assessment. Clinicians should document the app result as part of the history when it has influenced care seeking, then assess the lesion on its clinical merits.
A Belgian healthcare-sector decision-tree CEA compared SkinVision with usual care to the point of confirmed diagnosis and found lower modeled detection for melanoma (0.601 vs 0.874), BCC (0.585 vs 0.912), and SCC (0.706 vs 0.922), with lower pathway costs (€419 vs €536; €143 vs €145; €228 vs €233) as a consequence of missed detection rather than efficiency (Meertens et al., 2026). Even the melanoma best case still under-detected (0.834 vs 0.874). The authors do not support implementation or reimbursement in Belgium; this is a model to diagnosis, not lifetime ROI, and it is not a US coverage finding.
Patient Counseling Language
Short, consistent wording helps patients understand what an app or an AI-assisted examination can and cannot tell them. The following phrasing can be adapted to local practice.
When a patient asks whether to use a skin cancer app:
“These apps can help you learn about skin conditions, but they are not accurate enough to tell you that a spot is safe. If a mole is new, changing, bleeding, itching, or looks different from your other moles, please have it examined, whatever an app says.”
When a patient reports a “low risk” app result:
“Thank you for telling me. I will examine the spot the same way I would without the app. An app result does not replace an examination, and it should not decide whether you see a clinician about a changing or worrying lesion.”
When an AI-assisted device is used during the visit:
“This device gives me additional information about the lesion. It does not make the decision. I combine its result with your history, the examination, and dermoscopy to decide whether a biopsy or referral is needed.”
Teledermatology Is a Care Model, Not an AI Validation Shortcut
Teledermatology has evidence separate from AI. A 2026 systematic review and meta-analysis reported diagnostic concordance of 76% across skin conditions and 73% for skin cancers; dermoscopy improved skin-cancer concordance from 67% to 80% (Martyin et al., 2026). These findings support teledermatology as an access pathway in suitable settings. They do not establish that an AI classifier added to patient-submitted images improves diagnosis, referral efficiency, equity, or outcomes.
Any AI layer used with teledermatology should therefore be validated using the same image source, workflow, follow-up pathway, and patient population in which it will operate. A model trained on high-quality dermoscopy images should not be assumed to generalize to consumer smartphone photographs.
Longitudinal Inflammatory-Disease Monitoring
The most credible near-term expansion beyond lesion triage is measurement support for chronic inflammatory disease. Image-based systems are being developed to assess severity in atopic dermatitis, psoriasis, and acne, where repeated measures may support follow-up, treatment-response assessment, and research. The clinical question is whether a model improves decisions, burden, access, or outcomes; reproducing a score in a retrospective dataset does not answer it.
A systematic review of 45 studies and meta-analysis of 19 studies found pooled sensitivity of 80.5% and specificity of 96.2% for AI severity assessment, with performance varying materially by disease and scoring system (Cai et al., 2025). The included evidence was heterogeneous and was not sufficient to establish clinical benefit from automated severity scoring.
In a real-world observational study of patient-uploaded atopic dermatitis photographs, an AI-derived severity score correlated with clinician assessments, providing an early signal for longitudinal measurement rather than autonomous remote management (Okata-Karigane et al., 2025). The study’s single-country setting, image-quality exclusions, and small objective-reference subset limit generalizability.
For psoriasis, acne, and atopic dermatitis, prospective studies should compare AI-supported monitoring with standard care on treatment changes, patient-reported outcomes, clinician burden, access, and performance across skin tones and image conditions. Until then, AI severity scores should complement validated clinical instruments rather than replace them.
- What exact clinical decision does the tool support, and what is the reference standard?
- Does the validation population, image source, and workflow match local use?
- What are the group-specific sensitivity, specificity, calibration, and uncertainty estimates?
- What clinical action follows a positive or negative result, and who owns follow-up?
- What false-positive volume, specialist capacity, and patient burden should be expected?
- What prospective or real-world evidence shows benefit beyond image-level accuracy?
- How will drift, subgroup performance, and adverse events be monitored after deployment?
Regulatory status, integration, and contract:
- What is the FDA authorization (De Novo, 510(k), or PMA), its number, and the authorized intended use, users, and lesion criteria? Is the product instead marketed as a general wellness or informational tool outside device review?
- What imaging equipment, image-capture protocol, and staff training does the tool require, and how does its output reach the record and the referral pathway?
- What is the total cost of ownership, including devices, licenses, consumables, support, and staff time?
- How are liability, adverse-event reporting, and performance problems handled in the contract, and can the institution stop using the tool if local performance is inadequate?
Red Flags in Dermatology AI Procurement
- No subgroup performance by skin tone, image modality, anatomic site, and lesion subtype, or no account of how skin-tone labels were obtained
- Marketing that presents the tool as a replacement for dermatologist assessment or as a screening test when its authorization is narrower
- Validation only on curated or academic image sets, with no evaluation in the intended care setting and image source
- Diagnostic claims for a product without the corresponding FDA authorization
- Evidence that rests on early benchmark studies rather than external or prospective validation
- No ability to review or adjust the operating threshold against local referral capacity and clinical priorities
Clinical Bottom Line
Dermatology AI is clinically useful only when the task, users, images, and follow-up pathway are defined. A high score on a selected image dataset does not establish diagnostic safety, equitable performance, or patient benefit. The strongest current use cases are clinician-supervised adjuncts for defined lesion decisions and emerging longitudinal measurement support for inflammatory disease. Patient-facing informational tools and teledermatology can improve access and understanding, but neither should be presented as proof that automated self-triage is safe.
See Clinical AI Safety and Risk Management for common image-model failure modes, Physician AI Liability and Regulatory Compliance for documentation and oversight considerations, and Medical Ethics, Bias, and Health Equity for the clinical equity framework.
Questions About Dermatology AI
The Physician AI Handbook is free to read and cite, with no advertising. If this chapter informed your work, you can support the handbook.