| 摘 要: 3D医学图像分割是实现医学影像诊断、手术规划和治疗跟踪的前提与基础。腹部器官在影像上轮廓复杂、界限相对模糊,针对以上问题,提出了一种基于图卷积和有效自注意力的3D腹部器官分割网络。首先,在编码器端加入有效自注意力模块,有效地学习空间通道特征表示。其次,采用动态图卷积捕获腹部器官间的动态拓扑信息,同时有效突出腹部器官的特征。最后,在编码器端加入跳跃连接,融合不同分辨率的特征信息。实验结果表明,该方法在Amos22数据集上取得了较好的分割结果。 |
| 关键词: 深度学习;图卷积神经网络;注意力机制;医学图像分割 |
|
中图分类号: TP399
文献标识码: A
|
| 基金项目: 国家自然科学基金(62101497);浙江省基础公益研究计划项目(LTGY23F010001) |
|
| 3D Abdominal Organ Image Segmentation Method Based on Graph Convolutional Networks and Effective Self-Attention |
|
WANG Chuan, LI Yang, WEI Bo, JIANG Mingfeng
|
(School of Computer Science and Technology, Zhejiang Sci-Tech University, Hangzhou 310018, China)
wangchuan202403@163.com; yangli@zstu.edu.cn; weibo@zstu.edu.cn; m.jiang@zstu.edu.cn
|
| Abstract: 3D medical image segmentation is a prerequisite and foundation for medical imaging diagnosis,surgical planning, and treatment monitoring. Abdominal organs have complex contours and relatively ambiguous boundaries in images. To address these issues, a 3D abdominal organ segmentation network based on graph convolution and effective self-attention is proposed. First, an effective self-attention module is added to the encoder to effectively learn spatial-channel feature representations. Secondly, dynamic graph convolution is used to capture the dynamic topological information between abdominal organs and highlight the characteristics of abdominal organs. Finally, the skip connection is added to the encoder to fuse the feature information of different resolutions. Experimental results show that the proposed method achieves better segmentation results on the Amos22 dataset. |
| Keywords: deep learning; graph convolutional neural networks; attention mechanism; medical image segmentation |
