AI-assisted Dermatoscopic Imaging for Early Melanoma Detection

Thesis Abstract

The early and accurate detection of melanoma remains a critical challenge in dermatology, with disparities in diagnostic accuracy arising from variability in clinical expertise, access to specialty care, and the limitations of conventional dermatoscopic interpretation. This study addresses the problem by developing and validating an AI-assisted dermatoscopic imaging framework designed to improve sensitivity and specificity of melanoma detection in primary and secondary care settings. The aim is to create a robust, explainable AI pipeline that integrates high-resolution dermatoscopic images, standardized lesion segmentation, and clinician-in-the-loop decision support to facilitate early diagnosis and reduce unnecessary biopsies. Specific objectives are (1) to assemble a diverse dermatoscopic image database comprising 12,000 annotated lesions (6,000 melanoma, 6,000 benign nevi) from three dermatology centers; (2) to develop a multi-stage convolutional neural network (CNN) with attention mechanisms and a clinically oriented feature extractor to yield lesion classification with an area under the ROC curve (AUC) ? 0.93; (3) to implement explainable AI components (e.g., Grad-CAM, SHAP values) to provide lesion-region justifications aligned with the seven-point checklist and dermatoscopic criteria; (4) to evaluate model performance against pathologically confirmed diagnoses and compare it with senior dermatologist assessments; (5) to assess user acceptance, workflow integration, and potential impact on biopsy rates in a pilot implementation study. The methodology adopts a mixed-methods design combining retrospective image analysis and prospective user studies. The population includes patients presenting with pigmented skin lesions across three tertiary centers over five years, with a stratified sample of 8,000 images for training/validation and 4,000 images for independent testing; an additional 200 clinician participants participate in the pilot. Data collection instruments encompass standardized dermatoscopic imaging protocols, pathology reports as the reference standard, and structured questionnaires capturing clinician satisfaction and perceived utility. The data analysis plan employs a combination of deep learning model evaluation metrics (AUC, sensitivity, specificity, precision-recall, calibration curves), statistical comparison with dermatologist performance using McNemar’s test and DeLong’s test for correlated ROC curves, and explainability assessment through lesion-region highlight overlap with expert annotations (IoU, Dice) and clinician feedback analysis. A nested logistic regression model will examine factors predicting diagnostic improvement when AI assistance is used, incorporating physician experience, lesion morphology, and image quality. The study will address potential data bias through rigorous cross-site validation, gender and age stratification, and adversarial robustness checks. The theoretical framework situates the work within technology acceptance theory (TAM) and cue-utilization theory, complemented by the clinical decision support (CDS) affordance model to explain how interpretable AI outputs influence diagnostic behavior. A conceptual model links AI-derived probabilities, visual explanations, clinician judgments, and patient outcomes. Expected findings include high discriminative performance of the AI system (AUC ? 0.93), improved sensitivity in early-stage melanomas without a proportional rise in false positives, and positive clinician reception when explanations align with dermatoscopic criteria. The research anticipates demonstrating that AI-assisted interpretation reduces biopsy rates for benign lesions by a clinically meaningful margin while maintaining diagnostic safety. The study contributes to knowledge by delivering a validated, explainable AI dermatoscopic framework with real-world applicability, benchmarks for performance in diverse populations, and empirical evidence on workflow integration and user acceptance. The main conclusion is that AI-assisted dermatoscopic imaging can augment clinical decision-making for melanoma detection when accompanied by interpretable explanations and thoughtful integration into dermatology practice. Recommendations include scaling the imaging dataset across additional centers, refining the explainability module to reflect evolving diagnostic criteria, conducting longitudinal impact studies on patient outcomes, and establishing guidelines for regulatory and ethical governance of automated decision support in dermatology.