Skip to content

Clinical diagnosis of melanoma and dermoscopy

TL;DR — Dermoscopy is the established addition to naked-eye examination, with a relative diagnostic odds ratio of 15.6 (95% CI 2.9–83.7, P = .016) in clinical-setting meta-analysis, falling to 9.0 (1.5–54.6) after removing two outliers (Vestergaard 2008, PMID 18616769); the Cochrane review of 103 cohorts and 42,788 lesions confirms the benefit in referred populations and in experienced hands, and finds in-person evaluation more accurate than image-based (RDOR 4.6, 2.4–9.0, P < .001) while noting that primary-care data are sparse (Dinnes 2018, PMID 30521682). Formal algorithms perform worse prospectively than retrospectively: the 7-point checklist had 62% sensitivity in ten years of prospective surveillance against 78–95% in retrospective series, though with 97% specificity (Haenssle 2010, PMID 20226567). Reflectance confocal microscopy is the one adjunct with randomised evidence — in 3,165 dermoscopically suspicious lesions it raised positive predictive value from 18.9 to 33.3 and cut the number needed to excise by 43.4% (5.3 to 3.0), with all delayed melanoma diagnoses under 0.5 mm (Pellacani 2022, PMID 35648432). Convolutional neural networks match or beat dermatologists on curated image sets (pooled AUC 0.92, sensitivity 82%, specificity 87%) (Ye 2024, PMID 39088883), but no AI tool has been shown to improve a patient outcome in a prospective clinical workflow.

What clinical examination achieves before any instrument

  • Positive predictive value of a clinically suspicious lesion is low. Among 2,643 biopsies performed for suspicion of melanoma by 43 providers in one year at an academic centre, 165 were melanoma — PPV 6.4% (95% CI 5.5–7.4), i.e. 16 concerning lesions biopsied per melanoma found. Lesions >6 mm had PPV 11.5% (8.8–14.1) versus 2.6% (1.6–3.6) for smaller lesions, and older age (P < .001), male sex (P = .045) and non-trunk location (P < .001) predicted higher yield (Soltani-Arabshahi 2015, PMID 25582536).
  • The size criterion in the ABCD rule is supported, the algorithm less so. Applied bedside to 309 melanocytic tumours (46 melanomas, 263 nevi), the dermoscopic ABCD rule gave sensitivity 83% and specificity 45%, while the clinician's preliminary diagnosis gave comparable sensitivity (74%) with significantly higher specificity (91%); 19.6% of early melanomas were not preoperatively called malignant by either route (Ahnlide 2016, PMID 26351008).
  • Growth rate separates the phenotypes. In 404 consecutive invasive primary melanomas, median monthly growth was 0.12 mm for superficial spreading, 0.13 mm for lentigo maligna and 0.49 mm for nodular melanoma; one third of all melanomas grew ≥0.5 mm/month. Rapid growth tracked thickness (GMR 3.9 for 1.01–4 mm, 12.1 for >4 mm versus ≤1 mm) and mitotic rate (GMR up to 9.7 for >10/mm²), and occurred more often in men (GMR 1.7), people ≥70 (GMR 2.8), and — counter-intuitively — those with fewer nevi (<50, GMR 2.0) and fewer freckles (GMR 2.5) (Liu 2006, PMID 17178980). The patients least likely to be flagged as high-risk by nevus phenotype are disproportionately those with the fastest-growing tumours.
  • Nodular melanoma defeats morphological rules. Thin nodular melanoma often presents unremarkably clinically, and dermoscopic features are typically more suggestive of malignancy than clinical ones (Kalkhoran 2010, PMID 20231503) — the argument for instrumenting rather than eyeballing suspicious nodules.

Dermoscopy: the evidence and its limits

Comparison Estimate Source
Dermoscopy vs naked-eye, clinical setting (9 studies) Relative diagnostic OR 15.6 (2.9–83.7), P = .016; 9.0 (1.5–54.6) excluding two outliers Vestergaard 2008, PMID 18616769
Cochrane synthesis 103 cohorts, 42,788 lesions, 5,700 melanoma cases, 354 dermoscopy datasets Dinnes 2018, PMID 30521682
In-person vs image-based dermoscopy RDOR 4.6 (2.4–9.0), P < .001 favouring in-person PMID 30521682
In-person dermoscopy evidence base 26 evaluations, 23,169 lesions, 1,664 melanomas PMID 30521682
Visual inspection alone 13 evaluations, 6,740 lesions, 459 melanomas PMID 30521682

Cochrane's stated limitations are as important as its estimates: risk of bias was high or unclear for participant selection and flow, applicability concerns were high in three of four domains, and selective recruitment, non-reproducible diagnostic thresholds and missing observer-expertise detail were the recurring problems. Data supporting use in primary care are limited; formal algorithms are judged most useful for training and for less expert observers (PMID 30521682).

Algorithms in prospective use. Over ten years of prospective surveillance in 688 higher-risk patients, 127 melanomas were detected (50 in situ; mean invasive Breslow 0.57 mm). The 7-point checklist at its ≥3-point threshold flagged 79 of them — 62% sensitivity, against 78–95% reported retrospectively — but specificity was 97% versus 65–87% retrospectively. The remaining 48 melanomas scored <3 and were excised on complementary grounds such as lesional history or change on digital dermoscopy (PMID 20226567). In a reader study of 300 images by eight experienced dermatologists, pattern analysis correctly flagged 82% of melanomas and correctly spared 87.5% of monitored nevi, while the standard 7-point threshold recommended excision for 77.9% of melanomas and the revised threshold for 87.8%, at the cost of falling from 85.6% to 74.5% correct non-excision of monitored nevi (Argenziano 2011, PMID 21175563). Every algorithm trades sensitivity against unnecessary excisions on a fixed curve; none moves the curve.

Reflectance confocal microscopy

RCM is the only adjunct with a randomised trial and a specificity gain large enough to matter.

Setting Finding Source
Any lesion suspicious for melanoma (9 RCM datasets, 1,452 lesions, 370 melanomas) At fixed 90% sensitivity, specificity 82% for RCM vs 42% for dermoscopy; per 1,000 lesions at 30% melanoma prevalence, 280 fewer unnecessary excisions, 30 melanomas missed by both Dinnes 2018, PMID 30521681
Equivocal lesions (7 RCM datasets, 1,177 lesions, 180 melanomas) Specificity 86% RCM vs 49% dermoscopy; 296 fewer unnecessary excisions per 1,000 at 20% prevalence PMID 30521681
Randomised trial, 3,165 patients, 3 Italian referral centres, 2017–2019 PPV 18.9 → 33.3; benign:malignant ratio 3.7:1 → 1.8:1; number needed to excise 5.3 → 3.0 (−43.4%); all 15 delayed melanoma diagnoses were <0.5 mm thick Pellacani 2022, PMID 35648432
Amelanotic/hypomelanotic melanoma (7 studies, 1,111 lesions) Dermoscopy sensitivity 61% (37–81), specificity 90% (74–97); RCM sensitivity 67% (51–81), specificity 89% (86–92); RCM vs dermoscopy relative DOR 4.69 (0.81–27.3), P = .068 Lan 2020, PMID 31747045

Cochrane rates the RCM evidence base as at high or unclear risk of bias across almost all domains, with selective recruitment, unblinded reference standards and differential verification the specific problems, and calls for prospective real-world comparison (PMID 30521681). The 2022 randomised trial is that comparison for referral centres; its generalisability to primary care is untested. RCM is recommended by the European guideline "if available… in special cases" rather than routinely (Garbe 2025, PMID 39700658).

Artificial intelligence

Study Design Result
Haenssle 2018 (PMID 29846502) Inception v4 CNN vs 58 dermatologists (30 experts), 100-image test set Dermatologists level-I sensitivity 86.6% (±9.3), specificity 71.3% (±11.2); with clinical information specificity 75.7% (P < .05). CNN specificity 82.5% at matched sensitivities (P < .01 both levels); CNN ROC AUC 0.86 vs dermatologists' mean 0.79 (P < .01)
Brinker 2019 (PMID 31401469) CNN trained on 4,204 biopsy-proven images; 804 test images read by dermatologists from nine German university hospitals (19,296 recommendations) Dermatologists sensitivity 67.2% (62.6–71.7), specificity 62.2% (57.6–66.9); CNN sensitivity 82.3% (78.3–85.7), specificity 77.9% (73.8–81.8); all three McNemar tests P < .001
Ye 2024 (PMID 39088883) Systematic review/meta-analysis, 37 studies, 27 pooled Pooled sensitivity 82% (77–86), specificity 87% (84–90), AUC 0.92 (0.89–0.94). Head-to-head AUC 0.87 (DL) vs 0.83 (dermatologists); DL 0.90 vs junior 0.80 vs senior dermatologists 0.88; DL-assisted dermatologists 0.87 vs unassisted 0.76
Haggenmüller 2021 (PMID 34509059) Systematic review of 19 reader studies with direct human comparison 11 dermoscopic-image, 6 clinical-image, 2 whole-slide-histopathology studies
MacLellan 2021 (PMID 32289389) Prospective, 184 patients, 209 lesions, blinded dual pathology reference MelaFind sensitivity 82.5%/specificity 52.4%; Verisante Aura 21.4%/86.2%; FotoFinder Moleanalyzer Pro 88.1%/78.8%; teledermoscopist 84.5%/82.6%; local dermatologist 96.6%/32.2%

Two things follow. First, the human-versus-machine framing is the wrong one: the pooled data show assisted dermatologists (AUC 0.87) far above unassisted (0.76) and roughly equal to the algorithm alone (0.88) (PMID 39088883). Second, the prospective in-clinic result differs from the retrospective image-set result: in MacLellan's prospective series the local dermatologist had the highest sensitivity (96.6%) and by far the lowest specificity (32.2%), a trade-off no reader study on curated images reproduces (PMID 32289389). Randomised evidence that adding 3D total-body photography plus sequential digital dermoscopy without AI increased excisions without increasing melanoma detection (Soyer 2025, PMID 40136310) is the cautionary case for deploying image technology ahead of outcome evidence.

Teledermatology and triage

Teledermoscopy shows high sensitivity and specificity when dermoscopic images and expert interpretation are available, with reported reductions in Breslow thickness and waiting times and high satisfaction, though formal economic analyses are limited and AI results were mixed (López-Pardo Rico 2025, PMID 40940939). Real-world program data anchor what that means: among 1,571 lesions recommended for biopsy by MoleMap NZ dermoscopists with pathology available, 260 (17%) were melanoma — benign:malignant ratio 5.0:1, number needed to biopsy 6 — and among the 137 with thickness data, 92% were <0.8 mm with 74% in situ (Greenwald 2021, PMID 32114083). A high early-detection yield of this kind is simultaneously the case for and against the program, depending on how much of the in situ fraction is overdiagnosis (see screening and overdiagnosis).

Number needed to excise varies by an order of magnitude with setting and operator. In 118,668 pigmented lesions at one German academic department 2004–2013, overall NNT was 17.2 with a linear decrease toward older age (Schreieder 2024, PMID 39682200), against 6 in the teledermoscopy program (PMID 32114083) and 3.0 with adjunctive RCM in a referral centre (PMID 35648432). Any NNT figure is uninterpretable without the population and the operator.

Amelanotic and hypomelanotic melanoma

The recurring diagnostic failure mode. Amelanotic melanoma lacks the pigment on which every visual algorithm depends. Diagnostic accuracy is materially lower — dermoscopy sensitivity 61% and RCM sensitivity 67% for amelanotic/hypomelanotic lesions, against far higher figures for pigmented melanoma (PMID 31747045). Case-series work classifies the presentations as erythematous macule/patch on sun-exposed skin, dermal plaque or nodule without epidermal change, and exophytic nodule, and relates histological subtype to diagnostic delay (Gualandri 2009, PMID 19207640); comparative work in acral sites documents delayed diagnosis in amelanotic disease (Wu 2024, PMID 37690705). A 322-patient retrospective comparison found stage at diagnosis the strongest predictor of remission, progression and mortality in both amelanotic and melanotic groups, with no significant stage-stratified outcome difference — consistent with the harm being delay rather than intrinsic biology (Aviv 2026, PMID 42087524). See red flags and safety concerns.

Sites where the standard approach fails

Site Problem Evidence
Nail unit About two-thirds of nail-unit melanomas present as longitudinal melanonychia, which has a broad benign differential; diagnosis is often late and requires histopathology regardless of dermoscopic impression. Paediatric nail-unit melanoma is extremely rare and adult dermoscopic parameters have not been validated in children Conway 2023, PMID 36980308
Nail unit — the commonest mimic Across 90 subungual haemorrhage lesions in 64 patients, 84% showed more than one colour, 92% a homogeneous pattern, 42% globular patterns, 39% streaks, 54% peripheral fading and 16% nail-plate destruction — features overlapping with melanoma sufficiently that dermoscopy narrows rather than resolves the differential Mun 2013, PMID 23302009
Scalp and neck In 13,825 SEER head-and-neck melanomas 2011–2020, scalp/neck primaries had 5-year melanoma-specific survival 76.5% vs 82.7% elsewhere in the head and neck, adjusted HR 1.28 (95% CI 1.16–1.41), with higher odds of lung (aOR 2.39, 1.35–4.25) and multisite (aOR 2.23, 1.28–3.88) metastases Rashid 2026, PMID 41803580
Acral sites Amelanotic acral melanoma is diagnosed later than pigmented melanoma Wu 2024, PMID 37690705

Who detects the melanoma, and does timing matter?

Two findings sit uncomfortably together. In 816 consecutively diagnosed Italian melanoma patients, detection by a dermatologist — even incidentally — was the only factor that remained significantly associated with thinner tumours after adjustment, alongside univariate associations for female sex (OR for >1 mm 0.70, 0.50–0.97), higher education (0.44, 0.24–0.79), northern/central residence and skin self-examination (0.65, 0.45–0.93) (Carli 2003, PMID 12756097). But in a population-based telephone survey of 3,772 Queenslanders diagnosed with invasive melanoma 2000–2003 linked to registry pathology, there was no significant association between melanoma thickness and reported time to diagnosis for melanomas overall, for superficial spreading, or for nodular melanoma — the single exception being post-presentation delay in physician-detected nodular melanoma (Baade 2006, PMID 17116832).

The reconciliation is that who looks predicts thickness while how long it took does not — consistent with thickness at diagnosis being driven largely by tumour growth rate (PMID 17178980) and detection context rather than by patient or system delay. This weakens the standard "reduce delay to reduce thickness" argument for awareness campaigns without weakening the case for expert examination.

Biopsy technique and its consequences for staging

Question Evidence
Does partial biopsy misstage? In 323 referred cutaneous melanomas (median Breslow 0.54 mm), 76% were diagnosed by shave, 17% punch, 7% excision. Deep margin was positive in 33% of shaves and 23% of punches, versus 0% of excisions; residual melanoma at definitive excision in 40.6%, 60.4% and 19.0% respectively. Margin or sentinel-node recommendations changed in 6% of shave and 9% of punch patients, and there was no difference in unplanned reoperation by biopsy type (Jones 2023, PMID 36739830)
Is initial depth representative? Across 145 melanomas, 88% of non-excisional shave and punch biopsies gave Breslow depth ≥ the subsequent excision depth. Deep shave outperformed superficial shave and punch for melanomas <1 mm; excisional biopsy was the most accurate method (Ng 2003, PMID 12637923)

The practical reading is that excisional biopsy remains the reference and partial biopsy is acceptable in practice with an acknowledged understaging rate that changes management in roughly 6–9% of cases. Breslow thickness derived from a partial biopsy is a floor, not a measurement — see histopathology and prognostic factors and staging.

Interpretation rules for this page

  • Report the population before the accuracy figure. Dermoscopy's performance in referral populations does not transfer to primary care, and Cochrane says so explicitly (PMID 30521682).
  • Prospective algorithm performance is much worse than retrospective. The 7-point checklist loses ~20 sensitivity points moving from retrospective series to prospective surveillance (PMID 20226567).
  • Sensitivity and specificity trade on a fixed curve unless the modality changes. Only RCM has shifted specificity substantially at held sensitivity (PMID 30521681; PMID 35648432).
  • AI reader studies measure image classification, not care. No trial has reported AI improving a patient outcome; the one published randomised imaging intervention increased excisions (PMID 40136310). A recruiting 3,000-participant cluster-randomised primary-care study tests optional AI decision support, but its primary endpoint is diagnostic yield rather than morbidity or mortality (NCT06932172).
  • Number needed to excise is a property of the setting, not the technique — 17.2, 6 and 3.0 in three different settings here (PMID 39682200; PMID 32114083; PMID 35648432).
  • Amelanotic disease breaks the pigment-dependent tools and must be assessed separately (PMID 31747045).

Open questions

  • Does an AI-assisted workflow improve melanoma outcomes, or only reader-study accuracy and diagnostic yield? No prospective patient-outcome result was located as of 2026-09-01 (PMID 39088883; PMID 34509059; NCT06932172).
  • Does RCM's randomised 43.4% reduction in unnecessary excisions replicate outside high-volume referral centres (PMID 35648432)?
  • Can any tool improve amelanotic-melanoma sensitivity above the ~60–67% ceiling both dermoscopy and RCM currently show (PMID 31747045)?
  • Why do patients with fewer nevi and freckles have faster-growing melanomas, and should risk targeting account for it (PMID 17178980)?
  • What is dermoscopy's accuracy in unselected primary care? Cochrane identifies this as the principal gap (PMID 30521682).
  • Is a 74%-in situ detection profile in a teledermoscopy program a success or an overdiagnosis signal (PMID 32114083)?

References

  1. Vestergaard ME, et al. Dermoscopy compared with naked eye examination for the diagnosis of primary melanoma: a meta-analysis of studies performed in a clinical setting. The British journal of dermatology. 2008;159:669-76. PMID 18616769
  2. Dinnes J, et al. Dermoscopy, with and without visual inspection, for diagnosing melanoma in adults. The Cochrane database of systematic reviews. 2018;12:CD011902. PMID 30521682
  3. Haenssle HA, et al. Seven-point checklist for dermatoscopy: performance during 10 years of prospective surveillance of patients at increased melanoma risk. Journal of the American Academy of Dermatology. 2010;62:785-93. PMID 20226567
  4. Pellacani G, et al. Effect of Reflectance Confocal Microscopy for Suspect Lesions on Diagnostic Accuracy in Melanoma: A Randomized Clinical Trial. JAMA dermatology. 2022;158:754-761. PMID 35648432
  5. Ye Z, et al. Deep learning algorithms for melanoma detection using dermoscopic images: A systematic review and meta-analysis. Artificial intelligence in medicine. 2024;155:102934. PMID 39088883
  6. Soltani-Arabshahi R, et al. Predictive value of biopsy specimens suspicious for melanoma: support for 6-mm criterion in the ABCD rule. Journal of the American Academy of Dermatology. 2015;72:412-8. PMID 25582536
  7. Ahnlide I, et al. Validity of ABCD Rule of Dermoscopy in Clinical Practice. Acta dermato-venereologica. 2016;96:367-72. PMID 26351008
  8. Liu W, et al. Rate of growth in melanomas: characteristics and associations of rapidly growing melanomas. Archives of dermatology. 2006;142:1551-8. PMID 17178980
  9. Kalkhoran S, et al. Historical, clinical, and dermoscopic characteristics of thin nodular melanoma. Archives of dermatology. 2010;146:311-8. PMID 20231503
  10. Argenziano G, et al. Seven-point checklist of dermoscopy revisited. The British journal of dermatology. 2011;164:785-90. PMID 21175563
  11. Dinnes J, et al. Reflectance confocal microscopy for diagnosing cutaneous melanoma in adults. The Cochrane database of systematic reviews. 2018;12:CD013190. PMID 30521681
  12. Lan J, et al. The diagnostic accuracy of dermoscopy and reflectance confocal microscopy for amelanotic/hypomelanotic melanoma: a systematic review and meta-analysis. The British journal of dermatology. 2020;183:210-219. PMID 31747045
  13. Garbe C, et al. European consensus-based interdisciplinary guideline for melanoma. Part 1: Diagnostics - Update 2024. European journal of cancer (Oxford, England : 1990). 2025;215:115152. PMID 39700658
  14. Haenssle HA, 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 : official journal of the European Society for Medical Oncology. 2018;29:1836-1842. PMID 29846502
  15. Brinker TJ, et al. Deep neural networks are superior to dermatologists in melanoma image classification. European journal of cancer (Oxford, England : 1990). 2019;119:11-17. PMID 31401469
  16. Haggenmüller S, et al. Skin cancer classification via convolutional neural networks: systematic review of studies involving human experts. European journal of cancer (Oxford, England : 1990). 2021;156:202-216. PMID 34509059
  17. MacLellan AN, et al. The use of noninvasive imaging techniques in the diagnosis of melanoma: a prospective diagnostic accuracy study. Journal of the American Academy of Dermatology. 2021;85:353-359. PMID 32289389
  18. Soyer HP, et al. 3D Total-Body Photography in Patients at High Risk for Melanoma: A Randomized Clinical Trial. JAMA dermatology. 2025;161:472-481. PMID 40136310
  19. López-Pardo Rico M, et al. Teledermatology vs. Face-to-Face Dermatology for the Diagnosis of Melanoma: A Systematic Review. Cancers. 2025;17. PMID 40940939
  20. Greenwald E, et al. Real-world outcomes of melanoma surveillance using the MoleMap NZ telemedicine platform. Journal of the American Academy of Dermatology. 2021;85:596-603. PMID 32114083
  21. Schreieder L, et al. Impact of Patient's Age and Physician's Professional Background on the Number Needed to Treat in Malignant Melanoma Detection. Cancers. 2024;16. PMID 39682200
  22. Gualandri L, et al. Clinical features of 36 cases of amelanotic melanomas and considerations about the relationship between histologic subtypes and diagnostic delay. Journal of the European Academy of Dermatology and Venereology : JEADV. 2009;23:283-7. PMID 19207640
  23. Wu Q, et al. Clinicopathologic features, delayed diagnosis, and survival in amelanotic acral melanoma: A comparative study with pigmented melanoma. Journal of the American Academy of Dermatology. 2024;90:369-372. PMID 37690705
  24. Aviv B, et al. Comparison of Disease Progression Between Amelanotic Melanoma and Melanotic Melanoma. Pigment cell & melanoma research. 2026;39:e70083. PMID 42087524
  25. Conway J, et al. Adult and Pediatric Nail Unit Melanoma: Epidemiology, Diagnosis, and Treatment. Cells. 2023;12. PMID 36980308
  26. Mun JH, et al. Dermoscopy of subungual haemorrhage: its usefulness in differential diagnosis from nail-unit melanoma. The British journal of dermatology. 2013;168:1224-9. PMID 23302009
  27. Rashid S, et al. Prognostic Implications of Primary Site in Cutaneous Head and Neck Melanoma After the Implementation of Sentinel Node Biopsy: A SEER-Based Analysis (2011-2020). Annals of surgical oncology. 2026;33:5002-5008. PMID 41803580
  28. Carli P, et al. Dermatologist detection and skin self-examination are associated with thinner melanomas: results from a survey of the Italian Multidisciplinary Group on Melanoma. Archives of dermatology. 2003;139:607-12. PMID 12756097
  29. Baade PD, et al. The relationship between melanoma thickness and time to diagnosis in a large population-based study. Archives of dermatology. 2006;142:1422-7. PMID 17116832
  30. Jones S, et al. Clinical Impact and Accuracy of Shave Biopsy for Initial Diagnosis of Cutaneous Melanoma. The Journal of surgical research. 2023;286:35-40. PMID 36739830
  31. Ng PC, et al. Evaluating invasive cutaneous melanoma: is the initial biopsy representative of the final depth?. Journal of the American Academy of Dermatology. 2003;48:420-4. PMID 12637923