The early detection imperative: Melanoma is the most lethal form of skin cancer, and early-stage detection dramatically improves outcomes. Non-invasive imaging modalities - dermoscopy, Reflectance Confocal Microscopy (RCM), and Optical Coherence Tomography (OCT) - enable visualization of skin microstructures without biopsy, but their interpretation requires substantial training and expertise, creating variability in diagnostic accuracy.
Review objectives and scope: This 2023 systematic review from Oregon Health & Science University evaluated AI-based approaches applied to dermoscopy, OCT, and RCM for melanoma detection. The review adhered to PRISMA 2020 guidelines, searched Medline/PubMed, Cochrane, and Embase databases for publications between 2018 and 2022, and ultimately included 40 original peer-reviewed papers after rigorous eligibility screening.
Study composition: Of the 40 included papers, the vast majority (37) applied AI to dermoscopic images, with one paper using OCT and two using RCM. This distribution reflects the dominant use of dermoscopy in clinical dermatology practice and the availability of large public dermoscopy datasets (ISIC 2016, 2017, 2018, 2019, 2020; HAM10000) that have enabled widespread deep learning research.
Primary finding: All studies directly comparing AI performance with dermatologists reported equivalent or superior AI performance in melanoma detection. AI algorithms achieved mean sensitivity of 83.01% and mean specificity of 85.58% in dermoscopy-based comparative studies - with ROC values exceeding 80% in all head-to-head comparisons. OCT-based AI achieved 95% accuracy and RCM-based AI showed mean accuracy of 82.72%.
Search strategy: Comprehensive search terms covered 'melanoma', 'neural network', 'machine or deep learning', 'artificial intelligence', 'dermoscopy', 'reflectance confocal microscopy', and 'optical coherence tomography' across three databases. Initial screening of 287 articles was conducted by two independent reviewers; discrepancies were resolved by a third reviewer.
Inclusion and exclusion criteria: Studies were included if they directly compared AI-based evaluation of non-invasive imaging with human experts or histopathology, and reported diagnostic accuracy, AUC, or sensitivity/specificity. Studies were excluded if they did not directly diagnose melanoma, involved segmentation without classification, were editorials/commentary, or did not evaluate AI with one of the three specified imaging modalities.
Performance metrics analyzed: Three primary metrics were analyzed: accuracy (overall correct classification rate), sensitivity (true positive rate for melanoma detection), and specificity (true negative rate). For dermoscopy studies with direct human comparisons, mean sensitivity and specificity were calculated. For OCT and RCM studies, accuracy was used as the common metric. Secondary metrics including AUC, PPV, and NPV were also reported where available.
Metric appropriateness considerations: The review acknowledges that accuracy alone is insufficient for imbalanced datasets (where melanoma cases are fewer than benign lesions), as a classifier could achieve high accuracy simply by classifying all cases as benign. Therefore, sensitivity and specificity - which are independent of class prevalence - were prioritized for comparative analysis.
Consistent superiority over human readers: Multiple landmark studies confirmed AI superiority. In ISIC 2016, the top fusion algorithm achieved ROC of 0.86 versus dermatologist mean of 0.71 (p=0.001). In ISIC 2017, the best algorithm achieved ROC of 0.87 while dermatologists and residents achieved 0.66 (p<0.001), with the algorithm showing superior specificity (85.0% vs. 72.6%). One large study found that a deep learning model trained on 12,378 open-source images outperformed 136 of 157 dermatologists from 12 German university hospitals.
High-performing recent models: Several studies achieved outstanding results. Singh et al. reported accuracy exceeding 98.56% across four ISIC datasets using a neutrosophic features approach. Sayed et al. achieved 98.37% accuracy, 100% sensitivity, and AUC of 99% on ISIC 2020. The XceptionNet-based approach by Lu et al. achieved 100% detection accuracy with 94.05% sensitivity on HAM10000. The DERM (Deep Ensemble for Recognition of Melanoma) model achieved AUC of 0.93, sensitivity 85.0%, and specificity 85.3% on 7,102 dermoscopic images.
Vision transformers and novel architectures: SkinTrans (a vision transformer network) achieved 94.3% accuracy on HAM10000, demonstrating that transformer architectures can match CNN-based approaches. ZooME (Zoom-in Attention and Metadata Embedding) integrated patient demographics with image features, achieving 92.23% AUC on the large ISIC 2020 dataset of 33,126 dermoscopy images.
Subtype-specific performance: Winkler et al. evaluated CNN performance across melanoma localizations and subtypes. The model showed high sensitivity (>93.3%) for superficial spreading, nodular, and lentigo maligna melanomas, but lower sensitivity for acral lentiginous melanoma - an important finding because acral melanoma disproportionately affects darker-skinned populations where dermoscopy datasets are underrepresented.
Optical Coherence Tomography: Although only one OCT-specific paper met inclusion criteria, AI analysis of vibrational OCT (VOCT) images achieved 95% accuracy for skin cancer detection. OCT provides depth-resolved images up to 2mm below the skin surface using near-infrared light, offering a structural perspective complementary to dermoscopy's surface visualization. AI analysis of OCT images can distinguish tissue layers and structural disruptions characteristic of melanoma.
Reflectance Confocal Microscopy: Two RCM papers met inclusion criteria, showing mean accuracy of 82.72%. RCM uses a diode laser to capture high-resolution horizontal images at the cellular level as deep as the papillary dermis - the same depth accessed by surgical biopsy. AI analysis of RCM images can identify specific melanoma features including atypical melanocytes, pagetoid cells, and irregular dermal-epidermal junction patterns that are hallmarks of melanoma in situ.
Multi-modal potential: The different imaging depths and tissue information captured by dermoscopy (surface), OCT (structural, 2mm depth), and RCM (cellular, papillary dermis) suggest that combining AI analysis across modalities could provide complementary diagnostic information. No included study evaluated multi-modal AI integration, representing an important gap in current research.
Comparative performance context: The lower accuracy reported for RCM (82.72%) compared to dermoscopy top performers (>95%) may partly reflect the smaller available datasets for RCM training rather than inherent limitations of the modality. RCM image acquisition is also more technically demanding and less standardized across sites, making training data less consistent.
Primary care triage potential: Several studies explicitly evaluated AI deployment scenarios beyond specialist dermatology settings. The DERM model's performance (AUC 0.93) was contextualized against primary care physician AUC of 0.83 and dermatologist AUC of 0.91 from a meta-analysis, suggesting AI could provide dermatologist-grade recommendations in primary care. AI achieving 91.8% AUC from smartphone-captured images further supports point-of-care deployment.
Addressing the global dermatologist shortage: The review emphasizes that the benefits of AI are amplified in settings with limited access to specialized dermatological expertise. Regions with few dermatologists could deploy AI tools for initial screening of dermoscopic images, with positive cases flagged for specialist review - enabling earlier detection without requiring universal specialist access.
Standardization needs: A major challenge identified is variation in image quality, acquisition protocols, and preprocessing across studies. Without standardized image acquisition standards, AI systems trained on specific datasets may not generalize across different dermatoscopes, lighting conditions, or clinical workflows. The review calls for better standardization in image processing and consistent performance benchmarking.
Diversity and generalizability gaps: Most included studies used datasets dominated by fair-skinned patients. The lower CNN performance for acral lentiginous melanoma (more common in darker skin) highlights a critical gap. AI systems must be validated across diverse skin tones, ages, and geographic populations before claims of clinical equivalence with dermatologists can be broadly accepted.
Cross-population and skin type validation: The review explicitly identifies the need for AI validation studies in diverse populations, particularly darker skin tones and non-European ancestries where melanoma subtypes and dermoscopic presentation patterns may differ. Studies in Australia, sub-Saharan Africa, and Asia with different patient populations are needed to establish true generalizability.
Multi-modal AI integration: No reviewed study combined AI analysis across dermoscopy, OCT, and RCM. Given that each modality captures different tissue levels and characteristics, AI systems that synthesize information from multiple imaging modalities could provide more comprehensive and accurate diagnoses than any single modality approach. Future research should explore multi-modal fusion architectures.
Prospective clinical trials: The review underscores that all 40 included studies were retrospective in design, typically using already-diagnosed cases. Prospective studies where AI tools are deployed in real clinical workflows - measuring their impact on biopsy rates, missed diagnoses, and patient outcomes - are needed to establish clinical utility beyond diagnostic accuracy metrics alone.
Benchmarking standards: The inconsistency in reported metrics, dataset choices, and comparison methodologies across the 40 papers makes direct comparison difficult. Standardized benchmarking frameworks - similar to ISIC challenges - applied to OCT and RCM in addition to dermoscopy would enable more meaningful progress tracking in the field.