Scholarlink

<p>Breast cancer poses a persistent global health challenge, making early diagnosis indispensable for reducing mortality and improving patient prognosis. However, conventional detection paradigms are frequently impeded by the inherent complexity of lesions, characterized by minute dimensions, morphological heterogeneity, and indistinct boundaries. To address these impediments, this study proposes an advanced detection framework building upon YOLOv11n. We introduce three novel architectural components, specifically the C3k2-DCNv2-Dynamic, C2CGA, and CSFCN modules, designed to synergize feature extraction, fusion, and calibration. The C3k2-DCNv2-Dynamic module employs dynamic convolution and deformable mechanisms to robustly accommodate scale variations. Concurrently, the C2CGA module exploits a channel-guided attention mechanism within a multi-branch topology to heighten sensitivity toward complex lesion regions. Furthermore, the CSFCN module synthesizes contextual and spatial feature calibration to refine the identification of small targets. Extensive empirical evaluations validate the efficacy of the proposed method. The model achieved a precision of 66.6% and a mean Average Precision (mAP@0.5) of 86.2%, surpassing the baseline YOLOv11n by 5.5 and 2.4% points, respectively. Notably, detection accuracy for small G2-class lesions improved by a substantial 14.1% points. These findings substantiate the superior performance of our framework in resolving small and complex lesion detection, suggesting significant potential for clinical deployment.</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

Enhanced breast cancer detection framework based on YOLOv11n with multi-scale feature calibration

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年2月12日

查看全文
  • Ziqiong He,
  • Chen Zhang,
  • Chen Liang,
  • Wenyue Li

摘要

Breast cancer poses a persistent global health challenge, making early diagnosis indispensable for reducing mortality and improving patient prognosis. However, conventional detection paradigms are frequently impeded by the inherent complexity of lesions, characterized by minute dimensions, morphological heterogeneity, and indistinct boundaries. To address these impediments, this study proposes an advanced detection framework building upon YOLOv11n. We introduce three novel architectural components, specifically the C3k2-DCNv2-Dynamic, C2CGA, and CSFCN modules, designed to synergize feature extraction, fusion, and calibration. The C3k2-DCNv2-Dynamic module employs dynamic convolution and deformable mechanisms to robustly accommodate scale variations. Concurrently, the C2CGA module exploits a channel-guided attention mechanism within a multi-branch topology to heighten sensitivity toward complex lesion regions. Furthermore, the CSFCN module synthesizes contextual and spatial feature calibration to refine the identification of small targets. Extensive empirical evaluations validate the efficacy of the proposed method. The model achieved a precision of 66.6% and a mean Average Precision (mAP@0.5) of 86.2%, surpassing the baseline YOLOv11n by 5.5 and 2.4% points, respectively. Notably, detection accuracy for small G2-class lesions improved by a substantial 14.1% points. These findings substantiate the superior performance of our framework in resolving small and complex lesion detection, suggesting significant potential for clinical deployment.