Scholarlink

<p>Precise segmentation of pulmonary nodules in low-dose computed tomography is challenged by nodule heterogeneity, low contrast, and spatial overlap with adjacent anatomical structures. To address these issues, we propose CA-3DTransUNet, a segmentation framework based on the 3D-nnUNet architecture. The proposed network incorporates a Transformer 3D module in the bottleneck to model global volumetric dependencies and a CrossEMA3D module in the decoder to dynamically refine spatial features. Additionally, the wavelet transform is applied during the data preprocessing stage to augment input edge details. Evaluations on the LIDC-IDRI, LUNA16, and private BT datasets indicate the model’s performance. Specifically, on the LIDC-IDRI dataset, the model achieved a Dice Similarity Coefficient of 91.85 ± 0.43% [95% CI: 91.32–92.38], a Precision of 90.53 ± 0.51%, and a Sensitivity of 93.12 ± 0.42%. These results surpassed the hybrid architecture nnFormer, which attained a Dice score of 89.48 ± 0.52% (<i>p</i> = 0.014). These findings suggest that CA-3DTransUNet holds potential for the computer-aided analysis of pulmonary nodules.</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

CA-3DTransUNet with dynamic cross-scale fusion for pulmonary nodule segmentation

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月3日

查看全文
  • Kaikai Zhang,
  • Xiaowen Lan,
  • Yanhui Wang,
  • Lixin Wang,
  • Yuhan Liu,
  • Feng Guo

摘要

Precise segmentation of pulmonary nodules in low-dose computed tomography is challenged by nodule heterogeneity, low contrast, and spatial overlap with adjacent anatomical structures. To address these issues, we propose CA-3DTransUNet, a segmentation framework based on the 3D-nnUNet architecture. The proposed network incorporates a Transformer 3D module in the bottleneck to model global volumetric dependencies and a CrossEMA3D module in the decoder to dynamically refine spatial features. Additionally, the wavelet transform is applied during the data preprocessing stage to augment input edge details. Evaluations on the LIDC-IDRI, LUNA16, and private BT datasets indicate the model’s performance. Specifically, on the LIDC-IDRI dataset, the model achieved a Dice Similarity Coefficient of 91.85 ± 0.43% [95% CI: 91.32–92.38], a Precision of 90.53 ± 0.51%, and a Sensitivity of 93.12 ± 0.42%. These results surpassed the hybrid architecture nnFormer, which attained a Dice score of 89.48 ± 0.52% (p = 0.014). These findings suggest that CA-3DTransUNet holds potential for the computer-aided analysis of pulmonary nodules.