Development of a AI-assisted Dermoscopy Workflow for Melanoma Detection: Design, Implementation, Evaluation

Thesis Abstract

Early and accurate detection of melanoma remains a critical challenge in dermatology, where variability in lesion presentation and limited access to expert dermoscopy can delay diagnosis and treatment. This study addresses the problem by developing an AI-assisted dermoscopy workflow to enhance melanoma detection through integrated design, implementation, and evaluation of a decision-support system that augments clinician judgment with machine-generated analysis. The aim is to design a scalable, explainable, and clinically feasible pipeline that reliably assists dermatologists in screening pigmented lesions, reducing diagnostic uncertainty and potential misclassification. Specific objectives include (1) designing a modular dermoscopy pipeline that combines image pre-processing, lesion segmentation, feature extraction, and a hybrid classifier ensemble, (2) implementing interpretability mechanisms to ensure clinician trust through visual and textual explanations, (3) evaluating system performance against radiology and dermatology benchmarks using standardized datasets and prospective clinical validation, (4) assessing workflow integration, usability, and impact on diagnostic decision times, and (5) conducting a cost-benefit analysis to determine practical adoption pathways in diverse clinical settings. The methodology employs an explanatory sequential mixed-methods design anchored in the Technology Acceptance Model and the Diffusion of Innovations theory to examine both performance and adoption factors. The population comprises dermatology clinics across three metropolitan hospitals, with a total of 12 dermatologists participating and 1,200 dermoscopic images drawn from the publicly available ISIC Archive and hospital-specific repositories. The sample includes 600 images of histopathologically confirmed melanomas and 600 benign nevi matched for age, anatomic site, and imaging modality. Data collection instruments encompass a standardized dermoscopy imaging protocol, a software prototype integrating convolutional neural networks with traditional feature-based methods (e.g., color, asymmetry, border irregularity, dermoscopic structures), and a structured usability questionnaire. Inter-rater reliability for lesion labeling is assessed via Cohen’s kappa on a subset of 200 images reviewed by three expert dermatologists. The AI component employs a dual-path architecture a CNN-based lesion classifier trained with transfer learning on ImageNet-pretrained weights and fine-tuned on the dermoscopy dataset, and a feature-based radiomic-like descriptor analyzed via gradient boosting trees. Model interpretability is addressed through SHAP value explanations and Grad-CAM visualizations to highlight salient regions influencing the decision. The evaluation plan comprises internal validation with cross-validation (5-fold), external validation on an independent dataset of 300 images, and a prospective clinical validation involving 150 consecutive patient cases. Performance metrics include area under the receiver operating characteristic curve (AUC), sensitivity, specificity, positive predictive value, negative predictive value, and decision-curve analysis to quantify clinical usefulness. Statistical analyses employ DeLong’s test for AUC comparisons, paired t-tests for workflow time measurements, and multilevel logistic regression to account for clustering by dermatologist and site. A thematic analysis of qualitative data from clinician interviews (n=12) examines perceived usefulness, trust, and integration barriers, following Braun and Clarke’s methodology. Expected findings anticipate that the AI-assisted workflow will achieve an AUC ? 0.92 on external validation, with sensitivity and specificity values of 0.88 and 0.87, respectively, outperforming baseline dermoscopy assessments by a clinically meaningful margin. The explainable AI components are expected to correlate high-importance regions with established dermoscopic features, thereby enhancing clinician trust and acceptance. Usability assessments are projected to indicate efficient integration with existing workflows, with average decision-time reductions of 15–25% per case and a positive impact on diagnostic confidence. The study may identify site-specific factors influencing adoption, such as workflow redundancy, data quality, and training requirements, informing scalable deployment strategies. The anticipated contribution to knowledge includes empirically validated, interpretable AI-assisted dermoscopy workflows capable of enhancing melanoma detection accuracy while preserving clinician autonomy, along with a robust methodological blueprint for evaluating decision-support tools in dermatology. Theoretical contributions extend to empirical validation of technology acceptance and diffusion frameworks in medical AI, and practical implications include guidelines for data governance, model governance, and regulatory considerations. The study concludes that a carefully designed, explainable AI-assisted dermoscopy workflow can meaningfully improve early melanoma detection, reduce diagnostic variability, and be integrated into routine practice with manageable training and resource requirements; recommendations emphasize iterative clinician–AI co-design, continuous model updating, and multi-site prospective trials to confirm generalizability across populations and imaging systems.