摘要
为了更精确地对CT图像中的肝脏肿瘤边界进行分割,基于TransUNet分割网络,结合注意力模块(CBAM)以及混合注意力空洞空间金字塔池化模块(HA-ASPP),提出HA-TUNet级联分割网络,在提高卷积核感受野的同时,突出有用特征并抑制不重要特征,分割精度与肿瘤边缘的分割准确度优于改进前的TransUNet网络.基于LiTs公共数据集进行实验,HATUNet级联分割网络在肝脏与肿瘤分割中的Dice相似性系数指标较TransUNet网络分别提高了3.75%和3.39%,达到95.78%和73.35%,同时豪斯多夫距离95%相比TransUNet分别减少了0.56 mm和0.48 mm.
In order to more precisely segment the boundaries of liver tumors in CT images,the HA-TUNet cascaded segmentation network was proposed,based on TransUNet segmentation network,and incorporating the Convolutional Block Attention Module(CBAM)as well as the Hybrid Attention-Atrous Spatial Pyramid Pooling module(HA-ASPP).This network is designed to increase the receptive field while highlighting useful features and suppressing unimportant ones,achieving segmentation precision and tumor edge accuracy superior to the original TransUNet network.Experiments conducted on the LiTs public dataset show that the HA-TUNet cascaded segmentation network improves the DSC metric for liver and tumor segmentation by 3.75%and 3.39%respectively over the TransUNet network,reaching 95.78%and 73.35%.Additionally,the HD95 metric decreased by 0.56 mm and 0.48 mm respectively compared to TransUNet.
作者
李柯
刘文忠
秦镜淘
LI Ke;LIU Wenzhong;QIN Jingtao(College of Computer Science and Engineering,Sichuan University of Science and Engineering,Yibin,Sichuan 644000,China)
出处
《宜宾学院学报》
2024年第12期12-20,共9页
Journal of Yibin University
关键词
医学图像分割
CT图像
肝脏肿瘤分割
级联注意力网络
medical image segmentation
CT images
liver tumor segmentation
cascaded attention network
作者简介
第一作者:李柯(1998-),男,硕士研究生,研究方向为医学图像处理、数字图像处理;通信作者:刘文忠(1974-),男,讲师,研究方向为生物信息学、生物医学图像处理、计算生物学。
