Scholarlink

<p>Artificial intelligence has become a central methodological pillar in dermatological image analysis; however, the literature remains fragmented across classification, segmentation, detection, explainability, and clinical generalization. This review critically synthesizes 135 studies published between 2023 and 2026, with particular emphasis on convolutional neural networks (CNNs), Vision Transformers, hybrid CNN-Transformer architectures, ensemble systems, segmentation networks, detection frameworks, and explainable AI methods for skin lesion analysis. Rather than treating reported accuracy as a sufficient indicator of progress, the review evaluates the field through a broader set of diagnostic and morphological metrics, including accuracy, sensitivity, specificity, precision, Dice similarity coefficient, and intersection-over-union. This synthesis reveals a recurring evaluation problem, referred to here as the accuracy-sensitivity illusion, in which high aggregate accuracy may obscure clinically important sensitivity limitations and false-negative risks, particularly in melanoma-oriented classification. The reviewed evidence also indicates that segmentation performance is frequently reported under homogeneous benchmark conditions, whereas robustness across heterogeneous clinical images, different acquisition settings, and diverse skin phototypes remains insufficiently established. Hybrid architectures emerge as a promising direction because they can combine the local texture sensitivity of CNNs with the global contextual modelling capacity of attention-based mechanisms. Nevertheless, the field remains constrained by dataset concentration, limited external validation, inconsistent metric reporting, weak computational-cost analysis, and insufficient integration of explainability into clinically meaningful workflows. By integrating architectural taxonomy, dataset-level evidence, metric-based critique, and deployment-oriented limitations, this review provides a structured research agenda for developing more robust, interpretable, and clinically transferable AI systems for dermatological imaging.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Deep Learning for Dermatological Image Analysis: A Critical Survey of Architectures, Evaluation Metrics, Generalization, and Explainable AI

  • 期刊论文

  • 发表日期：

    2026年7月17日

查看全文
  • Ismail Kunduracioglu,
  • Ishak Pacal

摘要

Artificial intelligence has become a central methodological pillar in dermatological image analysis; however, the literature remains fragmented across classification, segmentation, detection, explainability, and clinical generalization. This review critically synthesizes 135 studies published between 2023 and 2026, with particular emphasis on convolutional neural networks (CNNs), Vision Transformers, hybrid CNN-Transformer architectures, ensemble systems, segmentation networks, detection frameworks, and explainable AI methods for skin lesion analysis. Rather than treating reported accuracy as a sufficient indicator of progress, the review evaluates the field through a broader set of diagnostic and morphological metrics, including accuracy, sensitivity, specificity, precision, Dice similarity coefficient, and intersection-over-union. This synthesis reveals a recurring evaluation problem, referred to here as the accuracy-sensitivity illusion, in which high aggregate accuracy may obscure clinically important sensitivity limitations and false-negative risks, particularly in melanoma-oriented classification. The reviewed evidence also indicates that segmentation performance is frequently reported under homogeneous benchmark conditions, whereas robustness across heterogeneous clinical images, different acquisition settings, and diverse skin phototypes remains insufficiently established. Hybrid architectures emerge as a promising direction because they can combine the local texture sensitivity of CNNs with the global contextual modelling capacity of attention-based mechanisms. Nevertheless, the field remains constrained by dataset concentration, limited external validation, inconsistent metric reporting, weak computational-cost analysis, and insufficient integration of explainability into clinically meaningful workflows. By integrating architectural taxonomy, dataset-level evidence, metric-based critique, and deployment-oriented limitations, this review provides a structured research agenda for developing more robust, interpretable, and clinically transferable AI systems for dermatological imaging.