Scholarlink

<p>Current methods for detecting pulmonary nodules in CT imaging often struggle to meet real-time requirements due to inefficient feature extraction. In addition, the small size and sparse distribution of the nodules lead to frequent missed or false detections. To address these issues, this study proposes DB-RTDETR, a real-time pulmonary nodule detection algorithm. Inspired by a bionic double-helix structure, the proposed Dual-Backbone network adopts two lightweight branches to achieve highly efficient and adaptive feature extraction. In addition, the Dynamic Range Histogram-Aware Self-Attention (DHSA) module dynamically adjusts local attention to refine feature structures, mitigating performance degradation and enhancing detection accuracy. Finally, the Adaptive Interactive Feature Fusion Module (AIFM) selectively emphasizes critical channel information and strengthens cross-feature interactions, significantly improving the detection precision for small and sparsely distributed nodules. Experiments conducted on the LUNA16 and LNDb datasets demonstrate that the proposed framework achieves an impressive 62.9% AP and 133 FPS on an RTX 3060 GPU, while reducing floating-point operations to only 26% of the baseline, surpassing state-of-the-art YOLO models in both speed and accuracy.</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

Real-time pulmonary nodule detection using a Dual-Backbone transformer framework

  • 期刊论文

  • 发表日期：

    2025年10月28日

查看全文
  • Kelei Sun,
  • Yihang Wang,
  • Huaping Zhou

摘要

Current methods for detecting pulmonary nodules in CT imaging often struggle to meet real-time requirements due to inefficient feature extraction. In addition, the small size and sparse distribution of the nodules lead to frequent missed or false detections. To address these issues, this study proposes DB-RTDETR, a real-time pulmonary nodule detection algorithm. Inspired by a bionic double-helix structure, the proposed Dual-Backbone network adopts two lightweight branches to achieve highly efficient and adaptive feature extraction. In addition, the Dynamic Range Histogram-Aware Self-Attention (DHSA) module dynamically adjusts local attention to refine feature structures, mitigating performance degradation and enhancing detection accuracy. Finally, the Adaptive Interactive Feature Fusion Module (AIFM) selectively emphasizes critical channel information and strengthens cross-feature interactions, significantly improving the detection precision for small and sparsely distributed nodules. Experiments conducted on the LUNA16 and LNDb datasets demonstrate that the proposed framework achieves an impressive 62.9% AP and 133 FPS on an RTX 3060 GPU, while reducing floating-point operations to only 26% of the baseline, surpassing state-of-the-art YOLO models in both speed and accuracy.