摘要：

用于生物图像和体积分割的过完备卷积结构很好地解决了传统编解码器方法不能精确分割边界区域的问题，但仍存在卷积运算不能较好学习全局和远程语义信息交互的缺点。对此，提出一种新的图像分割网络KTU-Net用于肝脏肿瘤的医学图像分割任务。该网络结构包括3个分支：1)学习捕捉输入细节和精确边缘的过完备卷积网络Kite-Net；2)学习高层特征的U-Net；3)学习输入体的序列表示并有效捕获全局多尺度信息的Transformer。设计包含早期融合和晚期融合2种融合方式的KTU-Net，采用一个混合损失函数来指导网络训练，使网络训练更加稳定。在LiTS肝脏肿瘤分割数据集上的实验结果表明，与先进的三维医学图像分割方法KiU-Net、TransBTS和UNETR相比，KTU-Net实现了更高或类似的分割精度。通过融合3个分支特征，肝脏肿瘤的平均Dice得分分别提高0.7%和2.1%，能够有效改善网络学习特征的质量，使肝脏肿瘤的分割结果更加准确，为医生判定准确的肝脏肿瘤细胞评估和治疗方案提供了可靠依据。

关键词: 计算机断层扫描, 肝脏肿瘤, 医学图像分割, 多分支, 特征融合

Abstract:

The overcomplete convolutional structure for biological images and volume segmentation is an excellent solution to the problem in which traditional codec methods cannot accurately segment the boundary regions. Although such methods perform well, the drawback that convolutional operations do not effectively learn global and remote semantic information interactions must be addressed. Accordingly, a new image segmentation network, KTU-Net, is proposed for the medical image segmentation of liver tumors. The network structure constitutes three branches: 1)Kite-Net, which is an overcomplete convolutional network that learns to capture input details and precise edges; 2)U-Net, which learns high-level features; 3)Transformer, which learns sequential representations of input bodies and efficiently captures global multiscale information. KTU-Net is designed for both early and late fusion, and a hybrid loss function is adopted to guide network training to achieve increased stability. From extensive experimental results regarding the LiTS liver tumor segmentation dataset, KTU-Net achieves higher or similar segmentation accuracy than other advanced 3D medical image segmentation methods such as KiU-Net, TransBTS, and UNETR. Fusing the three branching features, the average Dice scores of liver tumors are improved by 0.7% and 2.1%, achieving increased quality of features learned by the network and more accurate segmentation results of liver tumors, thus providing a reliable basis for doctors to perform precise liver tumor cell assessments and treatment plans.

Key words: CT, liver tumor, medical image segmentation, multi-branch, feature fusion