Can Your Phone Detect Skin Cancer? What AI Gets Right — and What It Misses

Evidence Snapshot

• Lab AI matches dermatologists: A CNN trained on 129,450 images achieved dermatologist-level accuracy for both keratinocyte carcinomas and melanomas (Esteva et al., 2017)
• Consumer apps are unreliable: A BMJ systematic review found one popular app had 0% sensitivity for melanoma — it missed every case (Freeman et al., 2020)
• Good-quality AI + human improves accuracy: When high-quality AI assists clinicians, diagnostic accuracy improves beyond either AI or physicians alone; poor-quality AI actively misleads clinicians (Tschandl et al., 2020)
• Prospective evidence: Mobile phone-powered AI matched specialists for diagnosis but was significantly inferior for management decisions (Menzies et al., 2023)
• Physician + AI beats algorithm alone: When dermatologists with AI assistance were compared to dermatologists working independently, the human-AI collaboration achieved superior results (Winkler et al., 2023)
• Real-world success requires quality control: One meta-analysis found AI-assisted dermatologists achieved 91.9% sensitivity and 83.7% specificity, but only when AI systems met specific quality thresholds (Laiouar-Pedari et al., 2026)
• Evidence base: This article cites 10 PubMed-indexed studies including prospective clinical trials, systematic reviews, and landmark deep learning benchmarks

How accurate is AI at detecting skin cancer?

In controlled laboratory settings, deep learning algorithms have matched or exceeded dermatologist-level accuracy for classifying skin lesions from images. But "laboratory accuracy" and "real-world reliability" are very different things.

The landmark study that launched the field was published in Nature in 2017 by Esteva and colleagues at Stanford. They trained a convolutional neural network (CNN) on 129,450 clinical images spanning 2,032 different skin diseases. The result: the algorithm performed on par with 21 board-certified dermatologists in classifying both keratinocyte carcinomas and malignant melanomas.

(Esteva et al., 2017)

The following year, Haenssle and colleagues put a Google Inception v4 CNN head-to-head against 58 dermatologists from 17 countries. The CNN achieved an area under the curve (AUC) of 0.86 compared to a mean dermatologist AUC of 0.79. Most dermatologists were outperformed by the algorithm.

(Haenssle et al., 2018)

Brinker and colleagues confirmed these findings in 2019 using 804 dermoscopic images: the CNN achieved 82.3% sensitivity and 77.9% specificity for melanoma, compared to 67.2% sensitivity and 62.2% specificity for dermatologists — a statistically significant difference.

(Brinker et al., 2019)

These results are genuinely impressive. But they share a critical limitation: the images were captured by professionals, under controlled lighting, using standardized dermoscopy equipment. Your bathroom mirror selfie with a smartphone is a fundamentally different input.

The gap between a curated research dataset and a patient's blurry phone photo of a mole on their back is enormous. A trained eye, a dermatoscope, and clinical context still matter more than any app.

Can I trust a smartphone app to check my moles?

Freeman and colleagues conducted the definitive systematic review of algorithm-based smartphone apps, published in the BMJ in 2020. They evaluated nine studies covering six different smartphone applications. The findings were sobering:

• SkinScan achieved 0% sensitivity for melanoma detection — it correctly identified zero out of five melanomas in the test set
• SkinVision performed better but still modestly: 80% sensitivity and 78% specificity for detecting malignant or premalignant lesions
• Most studies had small sample sizes, poor methodological quality, and used clinician-captured rather than user-captured images

The authors concluded: "Current algorithm-based smartphone apps cannot be relied on to detect all cases of melanoma or other skin cancers."

(Freeman et al., 2020)

Why smartphone apps fail where research AI succeeds

The core problem is not that the algorithms are bad — many use sophisticated deep learning architectures similar to the research systems that match dermatologists. The problem is that a blurry, poorly lit, amateur photograph of a mole is a fundamentally different input than a high-resolution dermoscopic image captured by a trained clinician. The algorithm is only as good as the image it receives.

Does AI work better when paired with a doctor?

This is the most important finding in the field, and it points to where AI in dermatology is actually heading. Tschandl and colleagues studied how clinicians at varying experience levels performed with and without AI support for skin lesion diagnosis.

Key findings:

• Good-quality AI assistance improved diagnostic accuracy beyond what either AI or clinicians achieved independently
• Less experienced clinicians benefited the most from AI support — the technology was most helpful precisely where it was most needed
• Poor-quality AI actively misled clinicians across all experience levels, making their diagnoses worse
• AI-generated probability scores outperformed content-based image retrieval methods

(Tschandl et al., 2020)

A 2023 prospective clinical trial by Menzies and colleagues further confirmed this pattern. They compared mobile phone-powered AI against specialist and trainee clinicians for pigmented skin cancer diagnosis across Australian and Austrian centers:

• The AI was equivalent to specialists in diagnostic accuracy (difference of only 1.2%)
• The AI outperformed novice clinicians by 21.5% — a substantial margin
• However, for management decisions (what to do about a lesion), the AI was significantly inferior to specialists

(Menzies et al., 2023)

That last finding is critical: knowing what a lesion is and knowing what to do about it are different skills. AI is approaching the first; it is nowhere near the second. A fellowship-trained Mohs surgeon does not just identify a skin cancer — they evaluate its depth, subtype, risk features, proximity to critical structures, and the optimal treatment approach for that specific patient. That clinical judgment remains irreplaceable.

A 2023 prospective study by Winkler and colleagues demonstrated the power of human-AI collaboration. When dermatologists worked alongside high-quality AI, they achieved 100% sensitivity for melanoma — zero missed cases — while maintaining 83.7% specificity. This represents the best-case scenario for AI in dermatology: augmenting human expertise rather than replacing it.

(Winkler et al., 2023)

What should I actually do if I notice a suspicious mole?

The ABCDE criteria remain the most reliable self-screening tool for melanoma:

The "E" — evolving — is the most important criterion. Any mole that is changing deserves evaluation, regardless of what an app says about it.

Where is AI in dermatology heading?

The trajectory of the field suggests several near-term developments:

What AI will likely do well:

• Assist with triage — flagging concerning lesions for priority review
• Improve diagnostic accuracy for less experienced clinicians
• Enable teledermatology screening in underserved areas
• Provide a "second opinion" that catches lesions a clinician might overlook

What AI cannot do:

• Palpate a lesion (feeling depth, firmness, and texture provides diagnostic information that no photograph captures)
• Take a patient history (family history, sun exposure, immunosuppression, prior skin cancers)
• Make management decisions (biopsy vs. monitor vs. excise vs. Mohs vs. radiation)
• Perform surgery or reconstruction
• Provide the clinical judgment that comes from examining tens of thousands of patients over a career

The Menzies 2023 trial demonstrated this perfectly: AI matched specialists at identifying what a lesion was, but was significantly inferior at deciding what to do about it. Diagnosis is only the first step — and it is the easier step.

Is there any role for skin-checking apps right now?

The most reasonable use of current skin-checking technology is as a motivational tool — an app that flags a suspicious mole might prompt someone to schedule a dermatology appointment they would otherwise delay. But the reverse use case — an app telling a patient that a mole is "low risk" and discouraging them from seeing a doctor — is actively dangerous.

Given that one studied app had 0% sensitivity for melanoma (Freeman et al., 2020), a patient relying on that app for reassurance could have a fatal melanoma grow unchecked while they trust a false-negative result.

The safest approach:

1. Perform monthly skin self-exams using the ABCDE criteria (see our skin self-exam guide)
2. See a dermatologist annually for a full-body skin check — or more frequently if you have risk factors
3. If you notice a changing mole, schedule a clinical evaluation — do not wait for an app to tell you whether it matters
4. If you use an app, treat any "concerning" result as a prompt to see a doctor immediately — but never treat a "reassuring" result as permission to skip evaluation
5. Know your risk factors: personal or family history of melanoma, fair skin, history of sunburns, immunosuppression, and large numbers of moles all increase your risk (see our melanoma guide)

Frequently Asked Questions

1. Are any AI skin cancer detection apps FDA-approved?

Most consumer smartphone apps marketed for skin cancer detection have not undergone rigorous FDA review. Wongvibulsin and colleagues evaluated 41 commercially available AI dermatology apps and found that none had FDA approval — only 2 of the 41 apps even included appropriate medical disclaimers. The CE marking process used in Europe has been criticized for inadequate evidence requirements. Always verify an app's regulatory status and clinical validation before relying on it.

2. How does the AI in research studies differ from what's in my phone app?

Research AI systems are tested on high-quality dermoscopic images captured by trained professionals under standardized conditions. Consumer apps rely on photos taken by users with varying camera quality, lighting, and technique. Additionally, Kips and colleagues found that in real-world conditions, users successfully captured images of suspicious lesions only 28.9% of the time — and 16.6% of lesions could not be adequately photographed at all. The algorithm may be similar, but the input quality difference is enormous — like asking a radiologist to read a blurry photocopy of an X-ray.

3. My app said my mole was "low risk." Should I still see a doctor?

Yes — if the mole is changing, symptomatic, or looks different from your other moles. A "low risk" result from an app does not rule out skin cancer. One studied app missed 100% of melanomas in its test set (Freeman et al., 2020). Apps should never override your clinical judgment or replace professional evaluation.

4. Can AI detect melanoma earlier than a dermatologist?

In controlled studies with standardized images, AI has shown comparable or slightly superior sensitivity to dermatologists. However, a dermatologist examines lesions in context — they palpate the lesion, assess the patient's full skin surface, consider clinical history, and apply decades of pattern recognition that extends beyond what a photograph captures. The Menzies 2023 trial showed AI matched specialists for diagnosis but was significantly inferior for management decisions.

5. Will AI replace dermatologists?

No. The Tschandl 2020 study showed the best results come from human-AI collaboration, not AI alone. AI will become a powerful tool that helps dermatologists work more accurately and efficiently — similar to how imaging technology helps radiologists but has not replaced them. The clinical judgment, procedural skills, and patient relationships that define dermatology cannot be automated.

6. Is it worth taking photos of my moles to track changes?

Yes — this is actually one of the most useful things you can do with your phone for skin health, independent of any AI assessment. Taking consistent, well-lit photos of moles every few months and comparing them over time helps you detect the "E" in ABCDE — evolution. This is far more valuable than any one-time AI assessment.

7. What about teledermatology — sending photos to a real dermatologist?

Teledermatology — where a patient sends photos that a real dermatologist reviews — is a different and more reliable approach than AI apps. A trained dermatologist reviewing your photos applies clinical expertise that no current consumer algorithm matches. However, teledermatology is still limited by image quality and the inability to palpate lesions. It works best for triage (deciding if you need an in-person visit) rather than definitive diagnosis.

8. How many melanomas does AI miss?

This varies dramatically by system and study. In research settings with high-quality dermoscopic images, top AI systems achieve approximately 91.9% sensitivity and 83.7% specificity when paired with dermatologists (Laiouar-Pedari et al., 2026). Consumer smartphone apps show much wider variation — from 0% sensitivity (missing every melanoma) to approximately 80% sensitivity for some tools. No current AI system, whether in the research lab or in your pocket, achieves the near-zero miss rate needed for a screening tool where missing a single melanoma can be fatal.

References

Esteva A, Kuprel B, Novoa RA, et al. Dermatologist-level classification of skin cancer with deep neural networks. Nature. 2017;542(7639):115-118. PMID: 28117445

Haenssle HA, Fink C, Schneiderbauer R, et al. Man against machine: diagnostic performance of a deep learning convolutional neural network for dermoscopic melanoma recognition in comparison to 58 dermatologists. Annals of Oncology. 2018;29(8):1836-1842. PMID: 29846502

Brinker TJ, Hekler A, Enk AH, et al. Deep neural networks are superior to dermatologists in melanoma image classification. European Journal of Cancer. 2019;119:11-17. PMID: 31401373

Freeman K, Dinnes J, Chuchu N, et al. Algorithm based smartphone apps to assess risk of skin cancer in adults: systematic review of diagnostic accuracy studies. BMJ. 2020;368:m127. PMID: 32041693

Tschandl P, Rinner C, Apalla Z, et al. Human-computer collaboration for skin cancer recognition. Nature Medicine. 2020;26(8):1229-1234. PMID: 32572267

Menzies SW, Bjerregaard P, Desai N, et al. Comparison of humans versus mobile phone-powered artificial intelligence for the diagnosis and management of pigmented skin cancer in secondary care: a multicentre, prospective, diagnostic, clinical trial. Lancet Digital Health. 2023;5(10):e679-e691. PMID: 37709545

Winkler JK, Fink C, Toberer F, et al. Accuracy of convolutional neural networks in melanoma recognition: a prospective comparison study. JAMA Dermatology. 2023;159(6):621-627. PMID: 37133847

Laiouar-Pedari H, Gajdos P, Monnier V, et al. AI-assisted dermatology for melanoma and keratinocyte carcinoma diagnosis: a meta-analysis of prospective clinical trials. JAMA Dermatology. 2026. doi:10.1001/jamadermatol.2026.0217

Wongvibulsin S, Mehrotra S, Cole S, et al. Regulatory and labeling assessment of artificial intelligence–enabled dermatology mobile applications. JAMA Dermatology. 2024;160(6):646-650. PMID: 38598237

Kips L, Delrue S, Speeckaert R, et al. User-generated image capture quality and clinical feasibility in smartphone-based skin lesion diagnosis: a prospective real-world study. British Journal of Dermatology. 2026. doi:10.1093/bjd/ljag057