Scholarlink

<p>Accurate and efficient pancreas segmentation serves as the cornerstone for quantitative diagnosis and precision treatment of pancreatic cancer. However, achieving high-precision segmentation from abdominal CT images remains challenging due to significant inter-individual variations in pancreatic morphology and interference from class imbalance and blurred boundaries. Although traditional U-shaped architectures and their transformer-integrated variants demonstrate strong performance in automated segmentation tasks, they often fail to effectively utilize intrinsic positional and channel features of medical images. To address pancreatic segmentation challenges, this study proposes a novel TriA-TransUnet network that integrates TriA-block which contains three task-specific attention modules with Transformers into the U-shaped architecture. Integrating this block into the encoder and skip connections enhance richer and more critical information features of the pancreas in image segmentation. Experimental results on two public datasets (NIH and MSD) reveal that our method achieves Dice Similarity Coefficients (DSC) of 85.51%and 80.91%, Intersection over Union (IoU) of 74.69% and 67.94%, and Precision (Pr) of 85.83%and 81.53%, respectively, significantly outperforming existing state-of-the-art approaches. Extensive evaluations demonstrate that the proposed framework excels in handling pancreatic morphological variations while maintaining computational efficiency, providing a robust solution for complex biomedical signal processing.</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

TriA-TransUNet: Triple attention-driven with morphological adaptation network for pancreas segmentation in CT images

  • 期刊论文

  • 发表日期：

    2025年11月10日

查看全文
  • Chong Luo,
  • Tiren Huang,
  • Xu Li

摘要

Accurate and efficient pancreas segmentation serves as the cornerstone for quantitative diagnosis and precision treatment of pancreatic cancer. However, achieving high-precision segmentation from abdominal CT images remains challenging due to significant inter-individual variations in pancreatic morphology and interference from class imbalance and blurred boundaries. Although traditional U-shaped architectures and their transformer-integrated variants demonstrate strong performance in automated segmentation tasks, they often fail to effectively utilize intrinsic positional and channel features of medical images. To address pancreatic segmentation challenges, this study proposes a novel TriA-TransUnet network that integrates TriA-block which contains three task-specific attention modules with Transformers into the U-shaped architecture. Integrating this block into the encoder and skip connections enhance richer and more critical information features of the pancreas in image segmentation. Experimental results on two public datasets (NIH and MSD) reveal that our method achieves Dice Similarity Coefficients (DSC) of 85.51%and 80.91%, Intersection over Union (IoU) of 74.69% and 67.94%, and Precision (Pr) of 85.83%and 81.53%, respectively, significantly outperforming existing state-of-the-art approaches. Extensive evaluations demonstrate that the proposed framework excels in handling pancreatic morphological variations while maintaining computational efficiency, providing a robust solution for complex biomedical signal processing.