Large Language Models in Dermatology: A Systematic Review of Clinical Accuracy and Image-Based Applications (Preprint)

BACKGROUND Recent advances in artificial intelligence have expanded its role in dermatology, particularly for image-based diagnostic support. While earlier efforts emphasized convolutional neural networks (CNNs) for image classification, newer models, such as transformers and large language models (LLMs), have enabled more integrated approaches that combine visual and textual data interpretation. Multimodal LLMs now offer capabilities such as contextual diagnosis and lesion analysis. While early results are encouraging, the accuracy, reliability, and practical application in real-world settings of these findings remain uncertain. This review summarizes current research on the use of AI-driven systems in dermatologic image analysis, highlighting their potential and limitations in clinical care. OBJECTIVE To systematically evaluate the accuracy, reliability, and clinical applicability of AI and multimodal LLM-based systems in dermatologic image analysis. METHODS This systematic review followed PRISMA-P guidelines. A structured search of PubMed and ScienceDirect (2020–2025) identified studies using the terms: “Artificial Intelligence” OR “ChatGPT” AND “Dermatology” AND “Image.” Studies were included if they evaluated AI or LLM-based methods for dermatologic image analysis. Exclusion criteria included non-original research, lack of image data, or non-dermatologic focus. Data were extracted on study design, population, study type, image modality, and reported diagnostic metrics. Statistical analysis included descriptive summaries and subgroup comparisons by model type and disease category. Where available, pooled sensitivity, specificity, and AUC were calculated using random-effects models. RESULTS 31 studies met the inclusion criteria, evaluating CNNs, transformer-based models, large language models (LLMs), and hybrid approaches in dermatologic image analysis. Models targeted melanoma, non-melanoma skin cancers (NMSC), benign dermatoses, and hair disorders. Across all studies, pooled mean sensitivity was 85.0% (95% CI: 80.0-90.0), specificity was 75.0% (95% CI: 65.0-85.0), and AUC was 0.90 (95% CI: 0.87-0.93). CNNs and transformers generally outperformed LLMs, especially in melanoma and NMSC detection. AI-assisted dermatologists demonstrated higher diagnostic sensitivity than unaided clinicians. Performance varied by disease type, image modality, and model class. CONCLUSIONS AI models, particularly CNNs and transformer-based systems, demonstrate strong diagnostic performance in dermatologic image analysis, especially for melanoma and non-melanoma skin cancers. CNNs and transformers consistently outperformed general-purpose LLMs and consumer apps, which remain limited in clinical reliability. AI support improved dermatologist sensitivity in a paired comparison. However, variability in study design, limited subgroup data, and underreporting of equity-related metrics highlight the need for more diverse, prospective evaluations. With rigorous validation and responsible integration, AI can augment dermatologic care and expand access globally.