基于条件卷积与注意力的肝脏分割算法  

Liver segmentation algorithm based on conditional parametric attention network

作　　者：赵浩辉 高永彬 杨淑群 胡小军 范应方[3] ZHAO Haohui;GAO Yongbin;YANG Shuqun;HU Xiaojun;FAN Yingfang(School of Electronic and Electrical Engineering,Shanghai University of Engineering Science,Shanghai 201620,China;Department of Hepatobiliary Surgery,the Fifth Affiliated Hospital of Southern Medical University,Guangzhou 510000,China;Hepatobiliary Department 1,Zhujiang Hospital of Southern Medical University,Guangzhou 510000,China)

机构地区：[1]上海工程技术大学电子电气工程学院,上海201620 [2]南方医科大学第五附属医院肝胆外科,广东广州510000 [3]南方医科大学珠江医院肝胆一科,广东广州510000

出　　处：《中国医学物理学杂志》2023年第6期701-708,共8页Chinese Journal of Medical Physics

基　　金：上海市“科技创新行动计划”社会发展科技攻关项目(21DZ1204900);广州市科技计划项目(202206010093)。

摘　　要：鉴于现有肝脏CT影像分割算法中存在的对比度较低、边界模糊、分割效果差问题,提出一种基于条件参数化卷积与注意力的分割网络(CPat-Net)。首先用条件参数化卷积替代残差网络中的常规卷积,其次将融合后的条件残差卷积模块集成至编码器中,用以提升模型容量和保持高效计算。然后利用特征注意(CPat)模块中的空间和通道注意力获取特征图的语义和细节信息,从而将局部特征与其全局依赖性更好地结合起来,最后利用深度监督进行多尺度语义信息的融合,提升方法的性能与鲁棒性。实验表明,在肝脏CT影像数据集中本文方法的Dice相似系数、交并比、Jaccrad系数分别达到了94.1%、90.3%、92.4%。相较于UNet、CENet、CSNet等前沿方法,本文方法在肝部分割上的准确度更为优异。In view of the low contrast,fuzzy boundary and poor segmentation results in the existing liver CT image segmentation algorithms,a conditional parametric attention network(CPat-Net) is presented.The method uses conditional parametric convolution to replace the conventional convolution in the residual network,and integrates the fused conditional residual convolution module into the encoder for improving model capacity and maintaining efficient computation.Then the spatial and channel attention mechanisms in the CPat module are used to obtain the semantic and detail information of the feature map,so as to better combine the local features with their global dependencies,and finally depth supervision is adopted to fuse multi-scale semantic information for improving the segmentation performance and robustness.The experiment reveals that the method has a Dice similarity cofficient,intersection over union and Jaccrad coefficient of 94.1%,90.3% and 92.4% on the liver CT image data set.Compared with the advanced methods such as UNet,CENet and CSNet,the proposed method has a higher accuracy in liver segmentation.

关 键 词：肝脏分割 卷积神经网络 条件参数化卷积 CPat-Net 

分 类 号：R318[医药卫生—生物医学工程] TP391[医药卫生—基础医学]