Robust and Explainable Convolutional Neural Network-Based Deep Learning for Automated Detection and Classification of Benign and Malignant Skin Cancer

Abstract

Early differentiation of benign and malignant skin lesions remains difficult because overlapping dermoscopic features, image artifacts, and uneven performance across clinical populations limit automated decision support. This research evaluates whether robust, explainable convolutional neural networks can improve lesion classification while supporting clinician trust and adoption. Guided by Trust in Automation Theory, the study adopts a quantitative, explanatory, retrospective cross-sectional, embedded multiple-case design combining image-level diagnostic comparisons with clinician-level assessments. The case framework comprises ISIC 2019 and an independent external dermoscopic cohort, with a target of 120 dermatologists and qualified clinicians across dermatology settings and specialties. From 25,331 ISIC images, 24,464 are identified as potentially eligible for binary benign-versus-malignant classification after excluding 867 actinic keratoses. An attention-enhanced EfficientNet-B3 model is compared with ResNet50, DenseNet121, and conventional EfficientNet-B3. Diagnostic accuracy, sensitivity, specificity, F1-score, AUROC, external generalization, robustness under four image perturbations, and Grad-CAM++ localization are evaluated separately from five-point Likert measures of explainability, reliability, robustness, usefulness, trust, and intended adoption. The analysis specifies paired model comparisons, descriptive statistics, scale reliability, Pearson correlations, and multivariable regression at p < .05. The manuscript reports 93.75% classification accuracy, 93.00% sensitivity, 94.00% specificity, and AUROC of .965 for the enhanced model; external validation reports AUROC of .925. Grad-CAM++ lesion overlap reaches .573 compared with .456 for conventional EfficientNet-B3. Clinician transparency scores rise from 3.30 to 3.97 when explanations are provided, while trust rises from 3.42 to 3.80. Regression models account for 52.0% of reported trust variation and 48.4% of adoption-intention variation. Correlations connect perceived diagnostic reliability with intended adoption (r = .604), and clinical trust with adoption (r = .579), while the adjusted explainability coefficient for trust does not meet significance (p = .091). These results highlight the importance of combining diagnostic discrimination, perturbation testing, explanation quality, and calibrated professional oversight before considering responsible clinical integration.