Scholarlink

Accurate classification of skin lesions into benign or malignant categories is crucial for early diagnosis and treatment of dermatological conditions. In this study, we present a comprehensive evaluation of Vision Transformer (ViT) models for binary classification tasks using a curated subset of the DermNet dataset. By leveraging a pre-trained ViT model fine-tuned on domain-specific data, we achieve a test accuracy of 94% with consistently high precision and recall scores. The results highlight the advantages of transformer-based architectures over traditional convolutional neural networks (CNNs), particularly in capturing global and contextual patterns in dermatological images. We describe the preprocessing pipeline, model architecture, and training strategies in detail, and we provide exhaustive evaluation metrics, confusion-matrix analysis, and interpretability visualizations. Finally, we discuss challenges and future directions—including computational efficiency, multi-class extensions, and clinical integration—showing the potential of ViTs to enhance dermatology decision-making.

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

Advancing Dermatology Diagnostics with Vision Transformers for Binary Skin Lesion Classification

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Magdalini Kreouzi,
  • Nikolaos Theodorakis,
  • Athanasios Anastasiou,
  • Konstantinos Kalodanis,
  • Aikaterini Sakagianni,
  • Effrosyni Bazakidou,
  • Iris Zoe Boufeas,
  • Georgios Dimitrakopoulos,
  • Konstantina Karathanasopoulou,
  • Maria Nikolaou,
  • Georgios Feretzakis

摘要

Accurate classification of skin lesions into benign or malignant categories is crucial for early diagnosis and treatment of dermatological conditions. In this study, we present a comprehensive evaluation of Vision Transformer (ViT) models for binary classification tasks using a curated subset of the DermNet dataset. By leveraging a pre-trained ViT model fine-tuned on domain-specific data, we achieve a test accuracy of 94% with consistently high precision and recall scores. The results highlight the advantages of transformer-based architectures over traditional convolutional neural networks (CNNs), particularly in capturing global and contextual patterns in dermatological images. We describe the preprocessing pipeline, model architecture, and training strategies in detail, and we provide exhaustive evaluation metrics, confusion-matrix analysis, and interpretability visualizations. Finally, we discuss challenges and future directions—including computational efficiency, multi-class extensions, and clinical integration—showing the potential of ViTs to enhance dermatology decision-making.