Scholarlink

<p>Accurate diagnosis and early detection of skin lesions significantly impact patient survival. Undetected or delayed diagnosis can allow skin lesions to progress to malignancy, resulting in severe health outcomes. To address these issues, we propose a deep learning-based computer-aided diagnosis (CADx) system for classifying skin lesions, optimized through a hybrid modeling strategy based on progressive defocusing guided attention (PDGA). This strategy combines the convolutional neural network (CNN) model’s ability to capture fine-grained local patterns with the transformer model’s capacity to model long-range dependencies and global context. To implement this integration, the transformer’s attention is initially guided by a lesion-specific mask generated from the CNN, and the guidance is progressively relaxed to enable broader contextual understanding. The effectiveness of the proposed PDGA in skin lesion classification was evaluated through comprehensive comparisons with both standalone models and other hybrid modeling approaches, including ensemble methods. As a result, PDGA achieved significantly higher performance than other methods, reaching an area under the curve (AUC) of 92.29% and a classification accuracy of 85.33% (<i>p</i> &lt; 0.001). This represents a 5.83% improvement over the baseline, along with lower computational inference costs. In summary, the proposed PDGA effectively combines the detailed feature extraction capabilities of CNNs with the global contextual understanding provided by transformer models, while also demonstrating significant potential as an assistive diagnostic tool in dermatology by improving both diagnostic accuracy and computational efficiency.</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

Progressive Defocusing Guided Attention in a Hybrid CNN-Transformer CADx System for Skin Lesion Classification

  • 期刊论文

  • 发表日期：

    2025年12月4日

查看全文
  • Hyeong-sik Won,
  • Jung-woo Chae,
  • Hyun-chong Cho

摘要

Accurate diagnosis and early detection of skin lesions significantly impact patient survival. Undetected or delayed diagnosis can allow skin lesions to progress to malignancy, resulting in severe health outcomes. To address these issues, we propose a deep learning-based computer-aided diagnosis (CADx) system for classifying skin lesions, optimized through a hybrid modeling strategy based on progressive defocusing guided attention (PDGA). This strategy combines the convolutional neural network (CNN) model’s ability to capture fine-grained local patterns with the transformer model’s capacity to model long-range dependencies and global context. To implement this integration, the transformer’s attention is initially guided by a lesion-specific mask generated from the CNN, and the guidance is progressively relaxed to enable broader contextual understanding. The effectiveness of the proposed PDGA in skin lesion classification was evaluated through comprehensive comparisons with both standalone models and other hybrid modeling approaches, including ensemble methods. As a result, PDGA achieved significantly higher performance than other methods, reaching an area under the curve (AUC) of 92.29% and a classification accuracy of 85.33% (p < 0.001). This represents a 5.83% improvement over the baseline, along with lower computational inference costs. In summary, the proposed PDGA effectively combines the detailed feature extraction capabilities of CNNs with the global contextual understanding provided by transformer models, while also demonstrating significant potential as an assistive diagnostic tool in dermatology by improving both diagnostic accuracy and computational efficiency.