Digital Dermoscopy AI for Early Skin Cancer Detection: Design, Implementation, Evaluation

Thesis Abstract

This study addresses the rising incidence of cutaneous melanoma and non-mmelanoma skin cancers and the limitations of conventional dermoscopic assessment in early detection, including variability in clinician interpretation and constrained access to expert dermatologists in primary care settings. The aim is to design, implement, and evaluate an AI-assisted digital dermoscopy workflow that enhances early cancer detection, triage accuracy, and decision support for clinicians. Specific objectives are (1) to develop a convolutional neural network (CNN)-based model capable of classifying dermoscopic images into malignant and benign categories with high sensitivity for melanoma; (2) to integrate the AI model into a clinician-facing dermoscopy platform with real-time decision support and explainable outputs; (3) to evaluate diagnostic performance against dermatologist consensus and histopathology as the reference standard across diverse imaging datasets; (4) to assess usability, workflow integration, and patient throughput implications in real-world clinical settings; and (5) to analyze potential biases and generalizability across skin types and imaging devices. A mixed-methods research design is employed, combining retrospective diagnostic performance evaluation with prospective usability testing and qualitative user feedback. The study uses a multicenter dataset comprising 25,000 dermoscopic images drawn from five tertiary dermatology centers, including a balanced representation of Fitzpatrick skin types I–VI, with histopathologically confirmed labels (melanoma, basal cell carcinoma, squamous cell carcinoma, and benign nevi). A stratified random sample of 8,000 images is reserved for model development and internal validation, while 2,000 images serve as an external test set to assess generalizability. For prospective evaluation, 300 consecutive patients presenting with pigmented skin lesions are recruited, and image acquisitions are performed by trained technicians using standard polarized dermoscopy. The AI model is developed using a CNN architecture with transfer learning from ImageNet and domain-specific fine-tuning, augmented by attention mechanisms and Grad-CAM-based explainability to highlight image regions contributing to decisions. Model performance is evaluated with sensitivity, specificity, area under the receiver operating characteristic curve (AUC), positive and negative predictive values, and decision curve analysis to determine potential clinical utility. Comparative performance against a panel of five board-certified dermatologists is analyzed using McNemar’s test and DeLong’s test for correlated AUCs. Calibration is assessed via Brier scores, and fairness across skin types is examined with subgroup analyses. In addition to diagnostic metrics, the study analyzes the impact of AI assistance on clinician performance through a randomized controlled simulation dermatologists interpret a curated set of 200 lesions with and without AI support, measuring changes in sensitivity, specificity, reading time, and confidence. Usability and workflow integration are evaluated using standardized tools, including the System Usability Scale (SUS) and task-technology fit surveys, complemented by semi-structured interviews exploring perceived trust, interpretability, and integration challenges. Ethical considerations address data privacy, informed consent for prospective participants, and governance of AI-generated recommendations. Expected findings include superior diagnostic performance of the AI-enabled workflow compared with unaided dermatologist interpretation, with an anticipated sensitivity of ?95% for melanoma and an AUC ?0.92 on external validation, while maintaining specificity >85%. External calibration is anticipated to show good concordance with actual outcomes, and Grad-CAM visualizations are expected to provide clinically meaningful localization supporting interpretability. The prospective usability study is expected to demonstrate improved reading efficiency with AI assistance, alongside high user acceptance (SUS scores >75) and favorable task-technology fit. The study also anticipates identifying residual biases related to device-specific image characteristics and underrepresented skin types, with recommendations for targeted data augmentation and ongoing model monitoring. The anticipated contribution to knowledge includes a robust, generalizable, and explainable AI-augmented dermoscopy framework that demonstrates feasibility and effectiveness for early skin cancer detection in diverse clinical settings, along with empirical evidence on workflow benefits, clinician acceptance, and ethical safeguards for deploying AI in dermatology. The study concludes that an integrated AI-assisted dermoscopy platform can augment diagnostic accuracy, reduce time-to-diagnosis, and support equitable care across populations when coupled with transparent explainability, rigorous validation, and continuous monitoring. Recommendations emphasize standardized imaging protocols, ongoing model updates with prospective performance surveillance, cross-institutional data sharing for fairness, and comprehensive clinician training to optimize adoption and patient outcomes.