Review scope and importance This Cochrane Diagnostic Test Accuracy review systematically evaluated computer-assisted diagnosis (CAD) systems for detecting melanoma, basal cell carcinoma (BCC), and cutaneous squamous cell carcinoma (cSCC) in adults. It included 42 studies covering 13,445 lesions with 2,452 confirmed cancers.
Types of CAD systems Two main technology categories were evaluated: digital dermoscopy-based CAD (Derm-CAD), which analyzes images from dermoscope-coupled cameras, and spectroscopy-based CAD (Spectro-CAD), which measures how light or electrical current passes through lesion tissue. Most Spectro-CAD studies used multispectral imaging (MSI); two used electrical impedance spectroscopy (EIS).
The diagnostic challenge Skin cancer diagnosis is high-stakes: missing a melanoma risks metastasis and death, while over-diagnosis causes unnecessary surgery, scarring, and patient anxiety. CAD systems aim to reduce both false negatives (missed cancers) and false positives (unnecessary excisions) across primary and specialist care settings.
Melanoma epidemiology Melanoma arises from uncontrolled melanocyte proliferation and can spread through lymphatic and blood vessels. It accounts for only a small fraction of skin cancers but up to 75% of skin cancer deaths. Five-year survival for stage I melanoma is 97-99%, dropping sharply to 32-93% for stage III disease depending on tumor thickness and lymph node involvement.
BCC and cSCC Basal cell carcinoma is the most common skin cancer and is usually localized but can cause significant disfigurement if untreated. Cutaneous squamous cell carcinoma can spread to regional lymph nodes and carries a 5-year metastatic survival rate of around 60%. Both are rising in incidence globally, driven by aging populations and cumulative UV exposure.
Clinical diagnostic pathway In the UK and similar systems, patients with suspicious lesions present first to general practitioners (GPs), who use visual inspection and the seven-point checklist before referring to specialists. Dermoscopy is the dominant non-invasive tool, enhancing accuracy especially for melanoma. CAD systems could potentially assist at multiple points: GP triage, remote teledermatology, and specialist second opinion.
CAD system architecture All CAD systems use machine learning trained on labeled lesion datasets. Classification algorithms include discriminant analysis, neural networks, support vector machines, and decision trees. Outputs range from binary malignant/benign results to risk scores and graphical lesion maps highlighting suspicious areas.
Comprehensive database search The review searched 10 databases from inception through August 2016, identifying studies of any design that evaluated CAD alone or compared it to dermoscopy in adults with lesions suspicious for melanoma, BCC, or cSCC. The reference standard was histological confirmation or clinical follow-up.
Data extraction and meta-analysis Two reviewers independently extracted data and assessed quality using QUADAS-2. Summary sensitivities and specificities were estimated separately for each CAD type using bivariate hierarchical models. Direct comparisons (studies with paired CAD and dermoscopy data) were supplemented with indirect comparisons using all available data.
Risk of bias concerns Study quality was generally poor: 34 of 42 studies raised concerns about patient selection, 19 did not clearly blind reference standard interpreters to CAD results, and 6 used differential verification (not all lesions biopsied). Many studies excluded difficult-to-diagnose lesions, artificially inflating apparent performance. Timing of tests was not specified in 28 studies.
Summary performance Pooled data from 22 Derm-CAD studies (8,992 lesions, 1,063 melanomas) showed sensitivity of 90.1% (95% CI: 84.0-94.0%) and specificity of 74.3% (95% CI: 63.6-82.7%). Sensitivity was consistent across studies, but specificity was more variable.
Clinical impact at 20% prevalence Applied to 1,000 lesions with 20% melanoma prevalence (typical for specialist clinic settings), Derm-CAD would correctly identify 180 melanomas but generate 206 false positives and miss 20 melanomas. This means more than half of positive Derm-CAD results would not actually be melanoma.
Comparison to dermoscopy Preliminary analysis of studies providing paired data found no statistically significant difference in sensitivity or specificity between Derm-CAD and dermoscopy performed by clinicians. This suggests Derm-CAD may offer comparable but not clearly superior performance to expert dermoscopy.
MSI-CAD summary performance Eight multispectral imaging studies (2,401 lesions, 286 melanomas) showed MSI-CAD sensitivity of 92.9% (95% CI: 83.7-97.1%) - slightly higher than Derm-CAD - but specificity of only 43.6% (95% CI: 24.8-64.5%), with very high heterogeneity between studies.
Clinical impact of low specificity At 20% melanoma prevalence, MSI-CAD would correctly identify 186 of 200 melanomas but generate 451 false positives. Nearly three-quarters of positive MSI-CAD results would be in patients without melanoma, representing a major overdiagnosis risk.
Electrical impedance spectroscopy (EIS) The Nevisense EIS system was evaluated in two studies. It measures cell characteristics by passing alternating current through skin lesions at 35 frequencies. Like MSI, it showed high sensitivity but low and variable specificity for melanoma.
Insufficient data for BCC and cSCC Only three studies addressed BCC and one addressed cSCC detection by CAD, providing insufficient evidence to draw any conclusions about CAD performance for keratinocyte cancers.
Highly selected populations limit generalizability Almost all studies enrolled narrowly defined populations, typically patients with lesions already selected for excision in specialist settings. This case-mix inflates apparent performance. CAD systems tested on the broad, unselected populations seen in primary care may perform substantially worse.
Three potential clinical roles CAD systems could fulfil three roles: (1) GP triage to improve referral appropriateness, requiring high sensitivity with acceptable specificity; (2) remote teledermatology triage, requiring high specificity to avoid overwhelming specialist clinics; and (3) expert second opinion in referral settings, requiring both high sensitivity and high specificity. Current evidence supports none of these roles definitively.
High sensitivity as safety net The strongest case for CAD is as a backup in specialist settings to minimize missed melanomas. A system with 90%+ sensitivity used alongside dermoscopy could reduce false negatives, particularly for difficult or atypical lesions. However, the cost in unnecessary excisions must be carefully weighed.
Evidence base weaknesses The key problems in existing CAD research are: poor study reporting, highly selected patient populations that do not represent real-world clinical settings, inconsistent algorithm thresholds, and lack of information about training set composition. Many studies were conducted at single institutions with non-representative lesion mixes.
Lack of primary care data Almost no studies evaluated CAD in unreferred (primary care) populations. Given that GPs and general practitioners represent the first clinical contact for most patients, this gap fundamentally limits understanding of where CAD could have the greatest population health impact.
Direct comparison studies needed The review found insufficient prospective studies directly comparing CAD to face-to-face dermoscopy in representative clinical populations. Such studies, with standardized protocols and representative patient populations, are needed to establish whether CAD adds value beyond current practice.
Continuous learning systems excluded This review did not evaluate AI systems that continuously update their algorithms as new cases are examined. Such systems may have very different performance characteristics from the fixed-algorithm systems studied and represent an important area for future evaluation.