AI-driven Mobile Dermoscopy for Early Melanoma Detection and Teledermatology

Thesis Abstract

Melanoma remains a leading cause of skin cancer mortality, with early detection markedly improving survival rates yet limited by access to specialist dermatology services and uneven clinical expertise in lesion assessment. This study addresses the problem by evaluating an AI-driven mobile dermoscopy platform designed to enable accurate, rapid, and scalable melanoma screening and to augment teledermatology workflows in both urban and rural settings. The aim is to determine (i) the diagnostic accuracy of a convolutional neural network–based dermoscopic image analysis pipeline for automated melanoma detection, (ii) the platform’s impact on wait times, referral appropriateness, and patient triage in teledermatology, and (iii) user and clinician acceptance, practical feasibility, and ethical implications of deploying mobile dermoscopy in routine care. Specific objectives include assessing sensitivity and specificity of the AI model against histopathology-verified diagnoses; evaluating changes in referral rates and time-to-diagnosis compared with conventional pathways; examining the interpretability of model outputs and the concordance with clinician judgments; exploring patient-reported outcomes related to usability, trust, and privacy; and identifying barriers and enablers to adoption across diverse populations. A mixed-methods design will be employed. The quantitative strand will recruit 1,200 participants presenting with pigmented lesions across five geographically diverse clinics over 12 months, with a stratified sampling approach to ensure representation by skin type, age, and lesion category. Data collection will involve standardized smartphone-based dermoscopy imaging, metadata capture (including lesion location, dermoscopic features, and clinical history), AI-generated diagnostic probabilities, and histopathology results for lesions proceeding to biopsy. The analytical framework will include receiver operating characteristic (ROC) analysis to estimate area under the curve (AUC), sensitivity, specificity, and likelihood ratios; calibration plots and decision-curve analysis to evaluate clinical usefulness; and time-to-diagnosis comparisons using survival analysis techniques. A multivariate logistic regression will examine predictors of accurate AI performance, and cost-effectiveness analysis will be conducted using incremental cost per correct diagnosis. The qualitative strand will involve semi-structured interviews with 25 dermatologists and 40 patients to explore perceived usefulness, trust in AI, workflow integration, and data privacy concerns, analysed via thematic analysis using a deductive framework anchored in Technology Acceptance Model (TAM) and Unified Theory of Acceptance and Use of Technology (UTAUT). The study will also apply the Explainable AI (XAI) paradigm to assess model interpretability and clinician–AI concordance. Ethical considerations include informed consent, data security in line with GDPR/HIPAA equivalents, de-identification protocols, and governance for remote diagnosis. The theoretical framework integrates the Health Belief Model to contextualize patient engagement, and the Technology Acceptance Model to interpret clinician adoption patterns, supplemented by the Theory of Planned Behavior for understanding user intentions. The AI model architecture comprises a transfer-learned CNN trained on a large, diverse dermoscopy dataset and fine-tuned with site-specific images; model outputs will include probability scores and attention maps to highlight salient dermoscopic features. The study anticipates key findings high diagnostic accuracy with AUC exceeding 0.92, improved triage efficiency with reduced average referral wait times by 25–40%, and substantial alignment between AI suggestions and dermatologist decisions in the majority of cases; robust usability scores indicating ease of use and high perceived reliability among patients and clinicians; and ethical acceptability with minimal reported privacy concerns when proper safeguards are implemented. The anticipated contribution to knowledge includes empirical evidence on the real-world performance and integration of AI-driven mobile dermoscopy within teledermatology, a framework for evaluating diagnostic confidence and explainability, and policy-relevant guidance on data governance and reimbursement. The study concludes that AI-assisted mobile dermoscopy can meaningfully shorten diagnostic pathways, enhance access to expert dermatology in underserved regions, and support accurate triage while preserving patient privacy. Recommendations emphasize standardized imaging protocols, continuous model monitoring for drift, ongoing clinician training in AI literacy, patient education on AI-assisted care, and the development of regulatory guidelines to govern deployment across heterogeneous healthcare settings.