Scholarlink

<p>The clinical diagnosis of dental fluorosis suffers from reliance on manual labor, high subjectivity, and insufficient efficiency. We propose a lightweight automated grading model based on deep separable convolutions and dual-axis attention. To address the scarcity of clinical annotation data, transfer learning is employed to enhance model generalization. An efficient convolutional architecture is designed to substantially reduce computational costs, enabling deployment in routine clinical settings. The proposed dual-axis attention mechanism mimics the diagnostic logic of observing tooth surface structures and lesions, improving recognition accuracy for subtle features. Experimental results demonstrate that while maintaining comparable performance to mainstream methods (80.00% accuracy, 79.88% macro F1 score), the model significantly reduces parameter count and computational load (by over 80%). This provides a feasible model for achieving efficient, objective clinical grading assistance for dental fluorosis.</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

LD2Net: real-time lightweight fluorosis grading with depthwise separable convolution and dual-axis attentional intelligence

  • 期刊论文

  • 发表日期：

    2026年4月6日

查看全文
  • Minghan Li,
  • Yun Wu,
  • Zhihao Li,
  • Chengdong Ye

摘要

The clinical diagnosis of dental fluorosis suffers from reliance on manual labor, high subjectivity, and insufficient efficiency. We propose a lightweight automated grading model based on deep separable convolutions and dual-axis attention. To address the scarcity of clinical annotation data, transfer learning is employed to enhance model generalization. An efficient convolutional architecture is designed to substantially reduce computational costs, enabling deployment in routine clinical settings. The proposed dual-axis attention mechanism mimics the diagnostic logic of observing tooth surface structures and lesions, improving recognition accuracy for subtle features. Experimental results demonstrate that while maintaining comparable performance to mainstream methods (80.00% accuracy, 79.88% macro F1 score), the model significantly reduces parameter count and computational load (by over 80%). This provides a feasible model for achieving efficient, objective clinical grading assistance for dental fluorosis.