Scholarlink

<p>Pneumonia remains a critical public health challenge, ranking as the third leading cause of death globally. Although deep learning has shown promise for segmenting pneumonia lesions in CT images, it still struggles to model long-range dependencies, effectively integrate multi-scale features, and cope with limited annotated data. These challenges hinder the extraction of fine-grained details and global contextual information of image regions, leading to blurred or incomplete segmentations, particularly for complex structures and lesion boundaries. To address these issues, a novel segmentation framework, Dual-Task Consistency regularized U-Net (DTCR-U-Net), is proposed, integrating a dual-task consistency regularization (DTCR) architecture with a multi-scale adaptive attention (MSAA) mechanism. Key innovations include: (1) task-level consistency regularization loss to jointly optimize segmentation and boundary detection tasks, ensuring geometric information alignment; (2) the MSAA mechanism, comprising Channel Contextual Enhancement (CCE) and Multi-Scale Global Spatial (MSGS) modules, to model inter-channel dependencies and cross-scale spatial correlations; (3) a multi-scale skip connection (MsSC) scheme to aggregate multi-level features during decoding; and (4) a semi-supervised learning paradigm to leverage unlabeled samples for enhanced generalizability with limited labeled data. Extensive experiments on the MosMedData + LIDC-IDRI, COVID-19-CT-Seg, and CC-CCII datasets demonstrate the superiority of DTCR-U-Net. On MosMedData + LIDC-IDRI, DTCR-U-Net achieves a Dice of 80.10%, F1 score of 83.76%, Sensitivity of 92.37%, Specificity of 99.87%, and Accuracy of 95.68%, outperforming all state-of-the-art baselines. Consistent improvements are observed on the CC-CCII benchmark (Dice 80.05%, F1 score 83.51%). Furthermore, under a semi-supervised setting on COVID-19-CT-Seg, DTCR-U-Net achieves the lowest Hausdorff Distance (4.947&#xa0;mm) and the highest Dice score (80.12%). These results demonstrate that DTCR-U-Net effectively tackles the long-range dependency modeling, multi-scale feature integration, and data scarcity, exhibiting strong robustness and promising clinical applicability for pneumonia lesion segmentation in CT images.</p> Graphical abstract <p>A novel segmentation framework, Dual-Task Consistency regularized U-Net (DTCR-U-Net), is proposed. DTCR network, is a dual-task learning framework specifically designed to address the challenges of limited labeled data and significant semantic gaps in pneumonia lesion segmentation. Traditional medical image segmentation model predominantly rely solely on a pixel-wise segmentation task to delineate infected regions, often neglecting critical structural information such as region boundaries and geometric features, resulting in suboptimal segmentation accuracy and blurred boundary delineation. To overcome this limitation, DTCR innovatively combines the pixel-level segmentation task (Task 1) with a boundary detection task (Task 2), optimizing both in a synergistic manner through shared representation learning.</p> <p></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

DTCR-U-Net: a dual-task consistency and multi-scale adaptive attention framework for pneumonia lesion segmentation

  • 期刊论文

  • 发表日期：

    2026年3月10日

查看全文
  • Jinping Liu,
  • Junqi Zhao,
  • Subo Gong,
  • Xiaohui Duan,
  • Guangyi Yang,
  • Jinxiu Li

摘要

Pneumonia remains a critical public health challenge, ranking as the third leading cause of death globally. Although deep learning has shown promise for segmenting pneumonia lesions in CT images, it still struggles to model long-range dependencies, effectively integrate multi-scale features, and cope with limited annotated data. These challenges hinder the extraction of fine-grained details and global contextual information of image regions, leading to blurred or incomplete segmentations, particularly for complex structures and lesion boundaries. To address these issues, a novel segmentation framework, Dual-Task Consistency regularized U-Net (DTCR-U-Net), is proposed, integrating a dual-task consistency regularization (DTCR) architecture with a multi-scale adaptive attention (MSAA) mechanism. Key innovations include: (1) task-level consistency regularization loss to jointly optimize segmentation and boundary detection tasks, ensuring geometric information alignment; (2) the MSAA mechanism, comprising Channel Contextual Enhancement (CCE) and Multi-Scale Global Spatial (MSGS) modules, to model inter-channel dependencies and cross-scale spatial correlations; (3) a multi-scale skip connection (MsSC) scheme to aggregate multi-level features during decoding; and (4) a semi-supervised learning paradigm to leverage unlabeled samples for enhanced generalizability with limited labeled data. Extensive experiments on the MosMedData + LIDC-IDRI, COVID-19-CT-Seg, and CC-CCII datasets demonstrate the superiority of DTCR-U-Net. On MosMedData + LIDC-IDRI, DTCR-U-Net achieves a Dice of 80.10%, F1 score of 83.76%, Sensitivity of 92.37%, Specificity of 99.87%, and Accuracy of 95.68%, outperforming all state-of-the-art baselines. Consistent improvements are observed on the CC-CCII benchmark (Dice 80.05%, F1 score 83.51%). Furthermore, under a semi-supervised setting on COVID-19-CT-Seg, DTCR-U-Net achieves the lowest Hausdorff Distance (4.947 mm) and the highest Dice score (80.12%). These results demonstrate that DTCR-U-Net effectively tackles the long-range dependency modeling, multi-scale feature integration, and data scarcity, exhibiting strong robustness and promising clinical applicability for pneumonia lesion segmentation in CT images.

Graphical abstract

A novel segmentation framework, Dual-Task Consistency regularized U-Net (DTCR-U-Net), is proposed. DTCR network, is a dual-task learning framework specifically designed to address the challenges of limited labeled data and significant semantic gaps in pneumonia lesion segmentation. Traditional medical image segmentation model predominantly rely solely on a pixel-wise segmentation task to delineate infected regions, often neglecting critical structural information such as region boundaries and geometric features, resulting in suboptimal segmentation accuracy and blurred boundary delineation. To overcome this limitation, DTCR innovatively combines the pixel-level segmentation task (Task 1) with a boundary detection task (Task 2), optimizing both in a synergistic manner through shared representation learning.