For a procurement committee, DermaSensor sits in two stacks of trust that should not be collapsed into one. The regulatory stack is clear: the device received FDA De Novo clearance in January 2024, was authorized for use by non-dermatologist physicians, and is indicated for patients aged 40 years or older.[1] The evidence stack is narrower: the 2025 DERM-SUCCESS evidence supports a real sensitivity gain in primary care decision-making, but it also shows a meaningful specificity penalty and leaves several implementation questions unanswered.[2]
That distinction matters because the search phrase “skin cancer treatment update 2025” points to a broader clinical area than this appraisal covers. This is not an update on immunotherapy, targeted therapy, radiation, surgery, or surveillance after treatment. It is an evidence appraisal of AI-assisted skin cancer detection in primary care, focused on whether DermaSensor’s peer-reviewed evidence justifies adding a new device-assisted step before referral decisions.
The short version is uncomfortable in the way many screening-adjacent technologies are uncomfortable: DERM-SUCCESS makes a credible case that the device can help primary care clinicians miss fewer malignancies, and that is not a trivial benefit. But the same evidence shows low standalone specificity, lower referral specificity after device assistance, and study conditions that are cleaner than routine primary care. A device can reduce false negatives and still push more uncertainty into an already constrained dermatology network.
What The Pivotal Evidence Actually Shows
The DERM-SUCCESS pivotal study was prospective and multi-site, enrolling 1,579 lesions from 1,005 patients across 22 U.S. sites. As a standalone device, DermaSensor reported 95.5% sensitivity for melanoma, basal cell carcinoma, and squamous cell carcinoma, with specificity of 20.7%.[2]
Those two numbers belong next to each other. A 95.5% sensitivity figure is the number that makes a primary care clinician lean forward; a missed melanoma is exactly the kind of failure decision support should try to prevent. A 20.7% specificity figure is the number that makes a referral manager, dermatologist, or access committee sit back down and ask what happens after the scan.
The companion clinical utility reader study is more relevant to workflow than standalone device performance because it asks what happens when PCPs see the device output. In that study, 108 PCPs reviewed 100 cases in a multi-reader, multi-case design. With device assistance, PCP management sensitivity improved from 82.0% to 91.4%, while referral specificity declined from 44.2% to 32.4%. The AUC improved from 0.708 to 0.762, and high-confidence decisions increased from 36.8% to 53.4%.[2]
| Measure | Without device | With device or standalone device | Procurement relevance |
|---|---|---|---|
| PCP management sensitivity | 82.0% | 91.4% | Fewer malignancies would be managed as non-referrals in the reader study. |
| PCP referral specificity | 44.2% | 32.4% | More benign lesions would be referred in the reader study. |
| AUC | 0.708 | 0.762 | Overall discrimination improved under study conditions. |
| High-confidence decisions | 36.8% | 53.4% | Clinicians felt more certain after seeing device output. |
| Standalone device sensitivity | Not applicable | 95.5% | The device was designed to favor cancer detection. |
| Standalone device specificity | Not applicable | 20.7% | The device alone generated many false positives. |
The strongest clinical argument for DermaSensor is the reduction in false-negative referral rate: PCP management false negatives fell from 18.0% to 8.6% with device assistance.[2] That is the part of the evidence that should not be waved away. If a primary care visit is the patient’s only realistic chance to have an evolving malignant lesion escalated, improving the sensitivity of that decision is meaningful.
But the specificity result is not a footnote. Referral specificity fell from 44.2% to 32.4%, and the reported p value for that decline was 0.0256.[2] In operational terms, the improvement in sensitivity was purchased by sending more benign lesions onward. The pivotal evidence therefore does not show that the device simply improves primary care skin cancer detection. It shows that, in the tested setting, the device shifts the miss-versus-over-referral balance toward fewer missed cancers and more benign referrals.
The Referral Burden Is Not A Theoretical Side Effect
Primary care skin lesion evaluation is not a closed-loop exercise. A scan result changes a clinician’s threshold, the clinician changes a referral decision, a dermatologist’s schedule absorbs the next step, and a patient waits. The clinical utility sub-analysis reported that, with device assistance, 254 malignancies were correctly referred and about 319 additional benign lesions were incorrectly referred.[2] That is not a reason to dismiss the device; it is a reason to stop describing the benefit as if the downstream work evaporates.
This is where prevalence does real damage to optimistic interpretation. DERM-SUCCESS used an enriched malignancy prevalence of about 50%, while real primary care is expected to have a much lower malignancy yield, with roughly 5% to 10% of biopsied lesions estimated to be malignant.[2] The study did not directly test DermaSensor in a low-prevalence, ordinary primary care lesion stream. Still, basic diagnostic arithmetic points in an unflattering direction: when prevalence falls, the positive predictive value of a low-specificity test falls with it.
That extrapolation should be labeled as extrapolation, not smuggled in as a study finding. DERM-SUCCESS did not prove that DermaSensor will overwhelm dermatology clinics. It also did not prove that extra benign referrals will be harmless. For a health system with long dermatology access times, the missing implementation evidence is not academic. It is the difference between “we caught more cancers” and “we caught more cancers while making access worse for other patients waiting to be seen.”
Confidence Increased, But Confidence Is Not Always Safety
The increase in high-confidence decisions from 36.8% to 53.4% is clinically interesting, especially for PCPs who do not evaluate suspicious lesions all day.[2] A short primary care visit often has too little time, too little dermatology backup, and too much room for uncertainty. A tool that gives a clinician enough confidence to refer a malignant lesion that might otherwise be watched deserves attention.
The same confidence signal also needs guardrails. A 2024 systematic review and meta-analysis of human-AI interaction in skin cancer diagnosis found that AI assistance improved accuracy for both dermatologists and non-dermatologists, but incorrect AI predictions harmed accuracy, with greater harm among non-specialists.[3] That finding lands directly on the primary care deployment question. The user group most likely to benefit from a nudge may also be the group most vulnerable to being nudged the wrong way.
The DERM-SUCCESS reader study supports the idea that device output can improve PCP decisions under test conditions. It does not settle how often a busy clinician will override a reassuring or alarming output when the lesion looks different in person, the patient history is concerning, the image is poor, or the appointment has already run over time. Procurement review should treat confidence as an implementation variable, not just a favorable usability result.
The Study Population Limits The Claim
The generalizability problem is large enough that it belongs in the center of the appraisal, not in a courtesy limitations paragraph. DERM-SUCCESS enrollment was 97.1% White, and Fitzpatrick skin types V and VI represented only 12.7% of participants.[2] For a device intended to support non-dermatologists across primary care, that is a thin basis for confidence in performance across skin tones.
The FDA recognized this gap by conditioning clearance on post-market performance testing in underrepresented populations.[1] As of this 2025 evidence appraisal, those post-market diversity results have not reported. That leaves a procurement committee with an awkward but familiar problem: the device is cleared, the pivotal evidence is peer-reviewed, and the population most represented in the trial is not broad enough to answer the equity question.
Underrepresentation is not just an ethics issue appended to model performance. It changes the evidentiary claim. A system cannot responsibly infer stable performance in groups that were sparsely represented, especially when the use case is primary care triage and the penalty for a false negative is delayed cancer evaluation.
The Test Conditions Were Cleaner Than Clinic
Several DERM-SUCCESS design choices make the evidence easier to interpret and harder to generalize. The pivotal study excluded crusted, ulcerated, or clinically infected lesions.[2] Those exclusions may be reasonable for device validation, but they also remove some of the messy lesions that make PCPs want decision support in the first place.
The clinical utility reader study used case review rather than an in-person exam, so PCPs reviewed images without palpation.[2] In real practice, touch is not ornamental. A clinician may feel induration, tenderness, friability, or other features that change suspicion. Removing tactile examination may understate usual PCP performance and inflate the apparent incremental value of image-plus-device review.
Rural representation also matters. Only 5% of PCPs in the utility study were from rural settings.[2] Rural clinics are often central to the access argument for AI triage, because dermatology is farther away and referral friction is higher. Yet the study offers limited direct evidence for the very settings where purchasers may feel the strongest pressure to adopt.
Sponsor involvement should slow the reading rather than end it. The DERM-SUCCESS clinical utility study was funded and sponsored by DermaSensor Inc.[2] Sponsored evidence can be valuable, and the study reports clinically relevant endpoints. But sponsorship increases the need to look closely at framing: enriched prevalence, excluded lesion types, reader-study conditions, and the choice to foreground sensitivity gains while the specificity cost carries the operational consequence.
This Pattern Did Not Start With DermaSensor
Historical skin cancer detection devices show a recurring regulatory and clinical pattern: very high sensitivity paired with low specificity. MelaFind was reported with 98.3% sensitivity and 9.9% specificity and was later discontinued; Nevisense has been reported with 97% sensitivity and about 31.3% specificity and remains marketed.[4] DermaSensor’s 95.5% standalone sensitivity and 20.7% specificity fit that broader pattern more than they overturn it.[2]
That context is useful because it prevents a false novelty story. The central question is not whether AI has finally made lesion triage frictionless. The question is whether a high-sensitivity, low-specificity device produces enough earlier cancer detection in the actual target workflow to justify the extra benign referrals it creates.
What A Health System Should Require Before Broad Deployment
DermaSensor has enough evidence to reject blanket dismissal. The sensitivity improvement in PCP management is clinically meaningful, and the false-negative reduction addresses a real primary care safety problem. A governance committee that treats the device as mere hype would be ignoring the strongest part of DERM-SUCCESS.
It does not have enough evidence to support an unqualified deployment claim. Clearance confirms that the device met the FDA’s De Novo pathway requirements for its authorized use; it does not prove net benefit in every primary care network, across all patient groups, or under ordinary low-prevalence clinic conditions.
- Real-world referral impact: whether benign dermatology referrals increase, by how much, and where the added demand lands.
- Post-market diversity performance: whether FDA-mandated testing in underrepresented populations confirms comparable sensitivity and specificity.
- Low-prevalence primary care performance: whether positive predictive value remains operationally tolerable when suspicious lesions are not artificially enriched.
- Workflow behavior: how often PCPs follow, override, or over-trust the device output during ordinary visits.
- Access consequences: whether earlier malignant referrals are achieved without worsening wait times for patients who also need dermatology evaluation.
Emerging foundation models may eventually change the technical ceiling. PanDerm has been described as trained on more than 2 million images, and newer multimodal AI systems have reported accuracy as high as 94.5% in research settings.[4] Those developments are worth watching, but they do not change today’s procurement decision for DermaSensor because they are not direct comparative implementation evidence and do not replace FDA-cleared device data.
For now, the defensible verdict is qualified. DERM-SUCCESS supports a sensitivity benefit for AI-assisted skin cancer detection in primary care. The net clinical value remains unproven until real-world implementation studies show what happens to referral volume, dermatology access, underrepresented patient groups, and ordinary low-prevalence clinic performance.
References
- Learnings from the first AI-enabled skin cancer device for primary care authorized by FDA, PMC, 2024.
- DERM-SUCCESS pivotal & clinical utility study, npj Digital Medicine, May 2025.
- Human-AI interaction systematic review & meta-analysis, npj Digital Medicine, 2024.
- The State of AI in Dermatology, Practical Dermatology, Jan/Feb 2025.