Scholarlink

<p>Early detection and precise fine-grained radiologic classification of lung cancer are crucial for formulating clinical treatment plans. However, existing computer-aided diagnosis (CAD) systems are mostly limited to binary nodule classification or treat radiologic feature classification and clinical TNM staging as isolated tasks. To address technical challenges such as drastic scale variations of tumors, subtle inter-class differences, and complex anatomical backgrounds, this paper proposes an end-to-end fine-grained lung cancer object detection algorithm based on an improved YOLO architecture. The proposed model accomplishes a complex 7-class detection task within a unified framework, encompassing normal tissue, three major non-small cell lung cancer (NSCLC) subtypes, and specific anatomical locations combined with TNM stages. Specifically, a Dilated Reparameterization Block (C3k2_DRB) is designed in the backbone to effectively enlarge the receptive field via multi-scale dilated convolutions, enhancing feature extraction for multi-scale tumors without introducing inference overhead. Meanwhile, an Explicit Positional Gating Optimization attention mechanism (C2PSA_EPGO) is introduced to adaptively focus on the most discriminative lesion textures and suppress irrelevant background noise. Furthermore, a dynamic upsampling strategy (DySample) is employed in the feature fusion stage to achieve content-aware spatial alignment, maximally preserving the morphological boundary details of the lesions. Experimental results on a complex lung cancer dataset comprising 1886 images demonstrate that the improved model achieves 97.6% Precision and 99.2% mean Average Precision (mAP@0.5) while maintaining extremely low parameters (2.46&#xa0;M) and computational cost (6.3 GFLOPs). Compared with mainstream object detection algorithms, the proposed method exhibits significant advantages in localization precision and fine-grained classification performance, demonstrating the potential of the proposed architectural enhancements for complex, fine-grained medical vision tasks.</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

Fine-grained lung cancer object detection via dilated reparameterization and explicit positional gating optimization

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月20日

查看全文
  • Yitao Wu,
  • Ziqiong He,
  • Su Zhang,
  • Chen Liang,
  • Chen Zhang,
  • Xue Yang,
  • Yizhe Zhang,
  • Shuwen Wang,
  • Kaiyu Cai

摘要

Early detection and precise fine-grained radiologic classification of lung cancer are crucial for formulating clinical treatment plans. However, existing computer-aided diagnosis (CAD) systems are mostly limited to binary nodule classification or treat radiologic feature classification and clinical TNM staging as isolated tasks. To address technical challenges such as drastic scale variations of tumors, subtle inter-class differences, and complex anatomical backgrounds, this paper proposes an end-to-end fine-grained lung cancer object detection algorithm based on an improved YOLO architecture. The proposed model accomplishes a complex 7-class detection task within a unified framework, encompassing normal tissue, three major non-small cell lung cancer (NSCLC) subtypes, and specific anatomical locations combined with TNM stages. Specifically, a Dilated Reparameterization Block (C3k2_DRB) is designed in the backbone to effectively enlarge the receptive field via multi-scale dilated convolutions, enhancing feature extraction for multi-scale tumors without introducing inference overhead. Meanwhile, an Explicit Positional Gating Optimization attention mechanism (C2PSA_EPGO) is introduced to adaptively focus on the most discriminative lesion textures and suppress irrelevant background noise. Furthermore, a dynamic upsampling strategy (DySample) is employed in the feature fusion stage to achieve content-aware spatial alignment, maximally preserving the morphological boundary details of the lesions. Experimental results on a complex lung cancer dataset comprising 1886 images demonstrate that the improved model achieves 97.6% Precision and 99.2% mean Average Precision (mAP@0.5) while maintaining extremely low parameters (2.46 M) and computational cost (6.3 GFLOPs). Compared with mainstream object detection algorithms, the proposed method exhibits significant advantages in localization precision and fine-grained classification performance, demonstrating the potential of the proposed architectural enhancements for complex, fine-grained medical vision tasks.