摘要
肺癌的及时诊断和治疗能够降低肺癌病人的死亡率,目前的主要诊断方法是采用计算机断层扫描(computer tomography,CT)技术,CT具有更高的分辨率和灵敏度,能够正确检测肺部中病灶位置。基于CT图像的肺实质准确分割是临床肺部疾病诊断的一个重要任务。针对肺实质分割时特征信息易丢失、前景和背景易混淆的问题,改进并提出了一种融合协调注意力和密集连接的空间卷积块的深度学习模型CAMCGU-Net(coordinate attention multi-level context gating U-Net)。主要改进如下:在编码器和解码器中间引入密集连接的空洞卷积块,帮助模型获取丰富多尺度特征信息,减少特征信息的丢失;在上采样后加入协调注意力(coordinate attention,CA)模块,高效整合空间坐标信息、增强目标对象的表示以提高模型对前景区域的定位能力,避免前景和背景混淆。在Kaggle肺分割数据集上的实验结果显示提出的模型得到的结果更加接近标注图像,在准确率(Accuracy,ACC)、特异性(Specificity,SP)、F1分数(F1-Score)等评估指标上均优于对比方法,能够更有效地分割肺实质。
The timely diagnosis and treatments of lung cancer can reduce the mortality of lung cancer patients. At present, the main diagnostic method is to use computer tomography technology. CT has higher resolution and sensitivity, and can correctly detect the location of lesions in the lung. Accurate segmentation of lung parenchyma based on CT images is an important task in clinical diagnosis of lung diseases. In view of the problems that feature information is easy to be lost and foreground and background are easy to be confused during lung parenchyma segmentation. This paper improves and proposes a deep learning model CAMCGU-Net(coordinate attention multi-level context gating U-Net)based on MCGU-Net model, which integrates coordinated attention and densely connected spatial convolution blocks. The main improvements are as follows: a dense connected hole convolution block is introduced between the encoder and decoder to help the model obtain rich multi-scale feature information and reduce the loss of feature information;After the up-sampling, a coordinate attention module is added to efficiently integrate the spatial coordinate information and enhance the representation of the target object to improve the positioning ability of the model to the foreground area and avoid confusion between the foreground and the background. The experimental results on Kaggle lung segmentation dataset show that the results of the proposed model are closer to the labeled images, and are better than the comparison method in accuracy, specificity, F1 score and other evaluation indicators, which indicating the proposed model can segment the lung parenchyma more effectively.
作者
杜佳成
余艳梅
汪恩惠
陶青川
Du Jiacheng;Yu Yanmei;Wang Enhui;Tao Qingchuan(College of Electronics and Information Engineering,Sichuan University,Chengdu 610065)
出处
《现代计算机》
2022年第24期52-56,共5页
Modern Computer
关键词
计算机断层扫描
肺实质分割
协调注意力
computed tomography
lung parenchyma segmentation
coordinate attention
