Scholarlink

<p>Lung cancer persists as a significant health challenge, emphasizing the necessity for efficient and accurate segmentation techniques for assisting early diagnosis, particularly for detecting intricate nodules that often lack clear visual boundaries. Computed Tomography (CT) images are used as the primary dataset because of their high-resolution imaging capabilities, which serve as a significant tool for detecting lung abnormalities. Existing attention-based models such as ECA (Efficient Channel Attention), SA (Shuffle Attention), and EGAM (Efficient Global Attention Mechanism) enhance feature representation, yet struggle with contextual understanding and multi-scale fusion. In this study, we propose a novel integration of RT-DETR into the YOLOv8 architecture, enriched with BiFPN for multi-scale learning, a DeepLabv3 + inspired decoder for semantic refinement, and atrous convolution for boundary preservation. The model’s loss formulation is carefully designed by combining Box Loss, Segmentation Loss, Classification Loss, and Distribution Focal Loss (DFL), all are dynamically weighted and optimized using Dung Beetle Optimization (DBO) technique. The proposed framework performed a thorough evaluation utilizing both the LUNA 16 dataset and clinical data collected from hospital sources. It attained superior results on all measures, comprising a Precision of 93.6%, a Recall of 92.1%, mAP of 92.9%, IoU of 91.5%, and an F1-score of 92.8%. The results underscore the model’s persistence in identifying small lesion variations and its computational proficiency in effectively integrating modular attention mechanisms with accurate segmentation. This makes it an auspicious approach for accurate and detailed lung nodule delineation.</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

A hybrid Attention-Transformer framework with Dung beetle optimization for Multi-Scale lung nodule segmentation in CT images

  • 期刊论文

  • 发表日期：

    2026年2月25日

查看全文
  • K. Vino Aishwarya,
  • A. Asuntha,
  • Jayanth Murugan

摘要

Lung cancer persists as a significant health challenge, emphasizing the necessity for efficient and accurate segmentation techniques for assisting early diagnosis, particularly for detecting intricate nodules that often lack clear visual boundaries. Computed Tomography (CT) images are used as the primary dataset because of their high-resolution imaging capabilities, which serve as a significant tool for detecting lung abnormalities. Existing attention-based models such as ECA (Efficient Channel Attention), SA (Shuffle Attention), and EGAM (Efficient Global Attention Mechanism) enhance feature representation, yet struggle with contextual understanding and multi-scale fusion. In this study, we propose a novel integration of RT-DETR into the YOLOv8 architecture, enriched with BiFPN for multi-scale learning, a DeepLabv3 + inspired decoder for semantic refinement, and atrous convolution for boundary preservation. The model’s loss formulation is carefully designed by combining Box Loss, Segmentation Loss, Classification Loss, and Distribution Focal Loss (DFL), all are dynamically weighted and optimized using Dung Beetle Optimization (DBO) technique. The proposed framework performed a thorough evaluation utilizing both the LUNA 16 dataset and clinical data collected from hospital sources. It attained superior results on all measures, comprising a Precision of 93.6%, a Recall of 92.1%, mAP of 92.9%, IoU of 91.5%, and an F1-score of 92.8%. The results underscore the model’s persistence in identifying small lesion variations and its computational proficiency in effectively integrating modular attention mechanisms with accurate segmentation. This makes it an auspicious approach for accurate and detailed lung nodule delineation.