AI-assisted smartphone dermoscopy for early melanoma detection in primary care

Thesis Abstract

Early melanoma detection remains a critical challenge in primary care, where non-specialist clinicians often encounter lesions with limited dermoscopic training and time constraints, leading to variable diagnostic accuracy and potential delays in referral to specialist care. This study investigates the feasibility, diagnostic accuracy, and clinical impact of an AI-assisted smartphone dermoscopy tool designed to aid primary care clinicians in identifying suspicious pigmented lesions indicative of melanoma, with the goal of enabling earlier detection and streamlined referral. The aim is to develop and validate a mobile AI-aided dermoscopy workflow that integrates image capture, automated lesion analysis, and decision support within primary care pathways. Specific objectives include (1) evaluating the diagnostic performance of the AI model against expert dermoscopy and histopathology references, (2) assessing the tool’s influence on clinician decision-making, referral appropriateness, and patient throughput, (3) identifying barriers and facilitators to adoption in diverse primary care settings, and (4) estimating potential health-economic impact through cost-effectiveness modeling. A mixed-methods design is employed in three sequential phases. Phase 1 comprises a multicenter diagnostic accuracy study across 12 primary care clinics, enrolling 1,200 patients presenting with pigmented skin lesions. High-resolution images are captured using standard smartphones equipped with the AI dermoscopy app, and outputs include a melanoma probability score and a recommended referral tier. The reference standard is histopathology for excised lesions and expert dermoscopic assessment for non-excised cases. Phase 2 uses a convergent parallel design to explore clinician experience and decision-making via semi-structured interviews with 20 clinicians and 200 observed consultations, analyzed through thematic analysis guided by the Technology Acceptance Model and Normalization Process Theory. Phase 3 integrates health-economic evaluation using a decision-analytic model to compare the AI-assisted pathway with standard care over a 5-year horizon, incorporating quality-adjusted life years and incremental cost-effectiveness ratios. Quantitative analyses in Phase 1 include calculation of sensitivity, specificity, positive and negative predictive values, area under the receiver operating characteristic curve (AUC), and calibration plots for the AI tool relative to histopathology and expert dermoscopy. Subgroup analyses examine lesion size, body site, Fitzpatrick skin type, and lesion concealment features. Multivariate logistic regression is applied to identify independent predictors of diagnostic discordance between the AI tool and reference standards. Phase 2 data are analyzed thematically, with analyst triangulation to ensure credibility and dependability. Phase 3 employs deterministic and probabilistic sensitivity analyses to test robustness of cost-effectiveness conclusions. The analytical framework is anchored in the Diffusion of Innovations theory and the Technology-Organization-Environment model to interpret adoption dynamics, supplemented by the Health Technology Assessment framework for the economic evaluation. Expected findings include (1) the AI-assisted dermoscopy tool achieving a pooled sensitivity above 85% and specificity above 80% for melanoma detection when deployed in primary care, (2) improved referral appropriateness with a measurable reduction in unnecessary specialist referrals by 15–20%, and (3) positive clinician attitudes toward usability and perceived impact on diagnostic confidence, tempered by concerns about workflow disruption and medico-legal implications. The qualitative component is anticipated to reveal facilitators such as rapid image capture, real-time feedback, and integration with electronic health records, alongside barriers including variability in device hardware, image quality, and patient consent processes. The economic evaluation is expected to demonstrate favorable cost-effectiveness under plausible uptake scenarios, particularly when paired with standardized referral pathways and clinician training. The study contributes to knowledge by providing robust, real-world evidence on AI-enabled dermoscopy as a scalable decision-support solution in primary care, detailing diagnostic performance, workflow integration, user acceptance, and health-economic implications. It informs policy on digital dermatology deployment, training curricula for primary care providers, and the design of integrated referral pathways to optimize melanoma outcomes. The main conclusion anticipates that AI-assisted smartphone dermoscopy can enhance early melanoma detection in primary care with acceptable accuracy, improved care efficiency, and favorable economic value, provided appropriate governance, ongoing clinician education, and rigorous data privacy safeguards are in place. Recommendations include standardized training modules for users, continuous model recalibration with diverse populations, explicit medico-legal guidelines, and implementation rollouts accompanied by monitoring metrics for diagnostic performance and patient outcomes.