Artificial intelligence-assisted analysis of clinical and dermoscopic skin lesion images: A comparative study with dermatologist evaluation

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Aim: Although dermoscopy has improved diagnostic accuracy for melanocytic lesions and melanoma in particular, inter-observer variability remains a clinical concern. This study evaluates the potential of an artificial intelligence (AI) ensemble to support dermatologists in a seven-class skin lesion classification task (including melanoma, basal cell carcinoma, actinic keratosis/intraepithelial carcinoma (AKIEC) and several benign lesions), using clinical and dermoscopic images. The comparison is intentionally restricted to an image-only, standardized classification setting and is not intended to represent full dermatologic diagnostic practice.Material and Methods: A dataset comprising more than 25,000 clinical and dermoscopic images obtained from the ISIC Archive and HAM10000 databases was used. A four-model ensemble deep learning architecture (ResNet-50, VGG-16, InceptionV3, and EfficientNet-B4) was developed using weighted voting aggregation. Dermatologists evaluated only the images and did not have access to patient age, sex, Fitzpatrick skin type, lesion site, duration, evolution, symptoms, prior treatments or palpation findings; the reader arm therefore reflects standardized image classification, not routine clinical practice. Model performance was compared with the dermatology panel on the same held-out image set using standard diagnostic metrics (sensitivity, specificity, F1, AUC-ROC). Grad-CAM was used to visualize model decision regions. Results: The AI ensemble achieved an overall seven-class accuracy of 94.2%, with an overall sensitivity of 93.8% (95% CI: 91.4-95.7) and specificity of 94.5% (95% CI: 92.3-96.2). The melanoma-specific sensitivity was 94.1%. On the same image-only test set, twelve board-certified dermatologists reached a mean accuracy of 88.5% (95% CI: 86.1-90.9). Ensemble modeling outperformed individual networks, and Grad-CAM outputs consistently mapped onto clinically relevant lesion regions.Conclusion: These findings should not be interpreted as AI being clinically superior to dermatologists, but as a comparative evaluation of standardized image-classification performance under an image-only setting. AI-assisted systems may serve as supportive tools in dermatology by assisting early detection, optimizing triage in high-volume settings and aiding teledermatology decision-making, complementing — not replacing — clinical expertise.