Scholarlink

<p>Multi-label chest X-ray classification, which aims to assign multiple pathological labels to images, is a core task and key support in thoracic disease diagnosis. However, image-based diagnosis remains challenging owing to heterogeneity in pathological size and distribution, coupled with similar visual characteristics and co-occurrence of distinct pathologies. This paper proposes HMCFNet, a novel dual-branch hierarchical multi-scale feature fusion network designed to integrate the complementary benefits of Mamba and CNN. Specifically, the network features a parallel dual-branch hierarchy composed of a Mamba branch for global representations and a CNN branch for local features. This hierarchical design enables efficient multi-scale feature extraction at various semantic levels with computational complexity that scales linearly with image size. Furthermore, we introduce a Hierarchical Global–Local Attention Fusion (HGAF) module that is equipped with spatial and channel attention to adaptively fuse semantic information between the two branches across different scales. Experimental results on the NIH ChestX-ray14, ChestX-ray11, and GXMC-CXR datasets show that HMCFNet matches or even surpasses state-of-the-art methods. This demonstrates its strong potential as a robust and generalizable solution for multi-label CXR classification in clinical practice.</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

HMCFNet: hierarchical Mamba-CNN fusion network for multi-label chest X-ray classification

  • 期刊论文

  • 发表日期：

    2026年3月9日

查看全文
  • Lianfeng Li,
  • Chengkun Li,
  • Yanhong Yang,
  • Yaning Mo,
  • Guodao Zhang,
  • Jinlian Che,
  • Yingfei Wang

摘要

Multi-label chest X-ray classification, which aims to assign multiple pathological labels to images, is a core task and key support in thoracic disease diagnosis. However, image-based diagnosis remains challenging owing to heterogeneity in pathological size and distribution, coupled with similar visual characteristics and co-occurrence of distinct pathologies. This paper proposes HMCFNet, a novel dual-branch hierarchical multi-scale feature fusion network designed to integrate the complementary benefits of Mamba and CNN. Specifically, the network features a parallel dual-branch hierarchy composed of a Mamba branch for global representations and a CNN branch for local features. This hierarchical design enables efficient multi-scale feature extraction at various semantic levels with computational complexity that scales linearly with image size. Furthermore, we introduce a Hierarchical Global–Local Attention Fusion (HGAF) module that is equipped with spatial and channel attention to adaptively fuse semantic information between the two branches across different scales. Experimental results on the NIH ChestX-ray14, ChestX-ray11, and GXMC-CXR datasets show that HMCFNet matches or even surpasses state-of-the-art methods. This demonstrates its strong potential as a robust and generalizable solution for multi-label CXR classification in clinical practice.