Scholarlink

<p>Accurate segmentation of lesions in lung CT images remains challenging due to blurred boundaries, small lesion sizes, and the scarcity of annotated data. To address these issues, this paper proposes a semi-supervised contrastive learning framework with a novel multiple attention UNet (MA-UNet) for lung CT image segmentation. The MA-UNet integrates a dual-attention module (DAM) and attention gates (AGs) to enhance spatial-channel feature refinement and boundary sensitivity. The DAM captures global context and channel-wise dependencies, while the AG emphasizes lesion-related features. Furthermore, residual blocks are used to improve gradient propagation and computational efficiency. To overcome limited annotations, we propose a contrastive learning framework that can fully utilize both labeled and unlabeled data to improve segmentation accuracy. To verify the validity of the methods and parameters design in this paper, we systematically carry out multiple ablation experiments. The experimental results show that the Dice, MIoU and Recall scores of MA-UNet based on comparative learning with only 1/2 ratio of labeled data are 78.41%, 88.78% and 91.79%, respectively, which are close to its supervised segmentation model, which effectively overcomes the problem of lack of labeled data.</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

Semi-supervised segmentation of lung CT images based on contrastive learning

  • 期刊论文

  • 发表日期：

    2025年5月12日

查看全文
  • Yiwen Qi,
  • Caibin Yao,
  • Hao Chen,
  • Xufei Wang

摘要

Accurate segmentation of lesions in lung CT images remains challenging due to blurred boundaries, small lesion sizes, and the scarcity of annotated data. To address these issues, this paper proposes a semi-supervised contrastive learning framework with a novel multiple attention UNet (MA-UNet) for lung CT image segmentation. The MA-UNet integrates a dual-attention module (DAM) and attention gates (AGs) to enhance spatial-channel feature refinement and boundary sensitivity. The DAM captures global context and channel-wise dependencies, while the AG emphasizes lesion-related features. Furthermore, residual blocks are used to improve gradient propagation and computational efficiency. To overcome limited annotations, we propose a contrastive learning framework that can fully utilize both labeled and unlabeled data to improve segmentation accuracy. To verify the validity of the methods and parameters design in this paper, we systematically carry out multiple ablation experiments. The experimental results show that the Dice, MIoU and Recall scores of MA-UNet based on comparative learning with only 1/2 ratio of labeled data are 78.41%, 88.78% and 91.79%, respectively, which are close to its supervised segmentation model, which effectively overcomes the problem of lack of labeled data.