基于MDRA-Net算法的肝脏CT图像分割方法  

Improved Liver Segmentation Method Based on MDRA-Net Algorithm

作　　者：李永进 LI Yongjin(Faculiy of Information Engineering and Automation,Kunming Universily of Science and Technology,Kunming 650500,China)

机构地区：[1]昆明理工大学信息工程与自动化学院,云南昆明650500

出　　处：《电视技术》2022年第6期58-62,78,共6页Video Engineering

基　　金：云南省智慧城市网络空间安全重点实验室开放课题项目(No.202105AG070010);调节肠代谢的绿色食品开发(No.202102AE090031)。

摘　　要：为了提高肝脏图像的分割精度,提出一种基于U-Net的改进网络即MDRA-Net算法。编码时使用多残差连接模块来提升网络的深度且加强肝脏特征的重复提取,解码时利用通道和空间注意力机制模块(scSE)来对模型提取后的特征进行筛选,加强有用特征传递,编码和解码都引入级联多尺度空洞残差模块用以充分捕捉肝脏图像的多尺度特征信息。在公开数据集LiTS17上验证了所提的算法,评价指标DICE的结果为98.689%,VOE结果为0.025,RVD结果为0.012。结果表明,与其他肝脏图像分割算法相比,本文提出的MDRA-Net算法在肝脏细节分割性能上有更好的表现。To improve the liver segmentation accuracy, this paper proposes an improved network based on U-Net, namely MDRA-Net.In encoding, the multi-residual connection module is used to improve the depth of the network and strengthen the repeated extraction of liver features. In decoding, the channel and spatial attention mechanism module(scSE) is used to screen the features after model extraction, and to strengthen useful feature transfer. In both encoding and decoding, cascaded multi-scale hole residual modules are introduced to fully capture the multi-scale feature information of the liver. The algorithm in this paper is verified on the public dataset LiTS17. The evaluation index DICE results are 98.689%, the VOE results are 0.025, and the RVD results are 0.012. The results show that compared with other liver segmentation algorithms, the proposed MDRA-Net algorithm has better performance in liver detail segmentation.

关 键 词：肝脏分割 残差模块 空洞残差模块 注意力机制 卷积神经网络 

分 类 号：TN911.73[电子电信—通信与信息系统]