基于U-Net改进的肺部轮廓与新冠病灶分割网络  

Improved Lung Contour and Lung COVID-19 Lesions Segmentation Network Based on U-Net

作　　者：林培阳 郑茜颖[1] LIN Peiyang;ZHENG Qianying(College of Physics and Information Engneering,Fuzhou University,Fuzhou 350108,China)

机构地区：[1]福州大学物理与信息工程学院,福建福州350108

出　　处：《电视技术》2023年第1期8-15,共8页Video Engineering

基　　金：福建省科技重点产业引导项目(2020H0007)。

摘　　要：针对肺部新冠病灶在医学成像中具有大小不均匀、多集中于肺部边缘且灰度与胸腔灰度相近的特点,提出一种基于U-Net改进的用于肺部轮廓分割和新冠病灶分割的网络模型。所提出的方法采用加深的编解码路径,使用带有残差连接的编码器子模块代替原始U-Net的标准卷积单元。为了提高高级特征的表征能力,在编码器和解码器中间加入自注意力机制,来学习特征的内在关系。整理一个用于分割训练的数据集,共2 973张新冠肺炎患者的肺部CT图片。实验结果表明,所提出的网络在肺部轮廓分割实验的Dice系数和F1系数分别达到了98.70%和98.89%,在新冠病灶分割实验中分别达到了87.47%和87.81%,优于其他对比模型。In view of the characteristics of the lung COVID-19 lesions in medical imaging, which are uneven in size, mostly concentrated on the lung edge, and whose gray level is similar to that of the chest, this paper proposes an improved lung contour and lung COVID-19lesions segmentation network based on U-Net. The proposed method adopts a deeper encoding and decoding path, and uses an encoder with residual connection to replace the standard convolution unit of the original U-Net. In order to improve the representation ability of high-level features, we add a self-attention mechanism between encoder and decoder to learn the intrinsic relationship of features.We collated a dataset for segmentation training, with 2 973 lung CT images of COVID-19 patients. The experimental results show that the Dice coefficient and F1 score of the proposed network reached 98.70% and 98.89% in the lung contour segmentation experiment,and 87.47% and 87.81% in the lung COVID-19 lesions segmentation experiment, respectively, which is superior to other comparison models.

关 键 词：深度学习 医学图像分割 U-Net 自注意力机制 

分 类 号：TP311.5[自动化与计算机技术—计算机软件与理论]