基于深度学习的腹部多器官图像分割  

Image segmentation of abdominal multiple organsbased on deep learning

作　　者：谢飞 权媚阳[3] 管子玉 段群[4] XIE Fei;QUAN Meiyang;GUAN Ziyu;DUAN Qun(School of Computer Science,Northwest Polytechnical University,Xi′an 710129,China;Academy of Advanced Interdisciplinary Research,,Xidian University,Xi′an 710126,China;School of Information Science and Technology,Northwest University,Xi′an 710127,China;School of Computer Science,Xianyang Normal University,Xianyang 712000,China)

机构地区：[1]西北工业大学计算机学院,陕西西安710129 [2]西安电子科技大学前沿交叉研究院,陕西西安710126 [3]西北大学信息科学与技术学院,陕西西安710127 [4]咸阳师范学院计算机学院,陕西咸阳712000

出　　处：《西北大学学报（自然科学版）》2021年第1期1-7,共7页Journal of Northwest University（Natural Science Edition）

基　　金：国家自然科学基金资助项目(61876145,61973249,61973250);陕西省教育厅服务地方科学研究计划资助项目(19JC041,19JC038)。

摘　　要：CT扫描是临床上腹部相关疾病诊断的常规检查方式,通过CT,医生能对腹部的器官结构和组织病变结构产生更加直观的观察,从而提高了疾病诊断的准确性,因此,精准地对CT图片进行图像分割有着非常重要的临床价值。传统的分割算法针对腹部形变较大、体积较小且组织边缘模糊的器官分割效果相对较差。为此,该文提出了基于改进nnUNet腹部多器官图像分割方法,在腹部CT图像上分割肝脏、胃、肠道和胰腺4个器官。该文利用自适应权重的损失函数对nnUNet网络进行改进,使得网络在分割过程中更加关注体积较小且样本数量相对较少的器官特征。实验表明,该文提出方法相对于现有传统的分割方法具有更高的准确性和敏感性。Clinically,CT scan is a routine examination method for the diagnosis of abdominal related diseases.Through CT,doctors can have a more intuitive observation of the organ structure and tissue pathological structure of the abdomen,thereby improving the accuracy of disease diagnosis.Therefore,accurate image segmentation of CT images has very important clinical value.However,traditional segmentation algorithms have relatively poor segmentation effects for organs with large abdominal deformation,small volume and blurry tissue edges.To this end,this paper proposes an improved nnUNet abdominal multi-organ image segmentation method to segment the liver,stomach,intestine and pancreas on the abdominal CT image.This paper proposes the use of adaptive softmax loss with adaptive weights to improve the nnUNet network,so that the network pays more attention to organ features with a small volume and a relatively small number of samples during the segmentation process.Relevant experiments show that the method proposed in this paper has higher accuracy and sensitivity than the existing segmentation methods.

关 键 词：腹部多器官分割 nnUNet 自适应权重损失函数 语义分割 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]