基于卷积神经网络的肺野分割和肺炎筛查  

Lung field segmentation and pneumonia screening using convolutional neural networks

作　　者：武卓越 田雪琴[2] 侯潇芮 崔磊[1] 田雪叶 王文晶 WU Zhuoyue;TIAN Xueqin;HOU Xiaorui;CUI Lei;TIAN Xueye;WANG Wenjing(School of Information Science and Technology, Northwest University, Xi′an 710127, China;School of Internet Application Technology, Shijiazhuang Institute of Technology, Shijiazhuang 050228, China;School of Mathematics, Northwest University, Xi′an 710127, China;Department of Obstetrics and Gynaecology, The First Affiliated Hospital of Xi′an Jiaotong University, Xi′an 710061, China)

机构地区：[1]西北大学信息科学与技术学院,陕西西安710127 [2]石家庄理工职业学院互联网应用技术学院,河北石家庄050228 [3]西北大学数学学院,陕西西安710127 [4]西安交通大学第一附属医院妇产科,陕西西安710061

出　　处：《西北大学学报（自然科学版）》2022年第4期571-580,共10页Journal of Northwest University（Natural Science Edition）

基　　金：陕西省自然科学基金(2021JQ-461);西安交通大学医学院第一附属医院科研发展基金(2021ZYTS-07)。

摘　　要：由于X光胸片影像受到各类组织阴影及病灶的影响,目前肺野分割算法的结果往往存在空洞或者边缘不光滑等问题。针对此类问题,该文提出了一种基于多尺度卷积和特征金字塔的肺野分割网络,此网络利用多尺度卷积模块和多尺度特征融合模块提取和融合多尺度特征,在JSRT数据集上PA和Dice指标分别达到98.76%和97.94%,在Montgomery数据集上PA和Dice指标分别达到了98.96%和97.85%。该文将肺野分割网络进一步应用到肺炎筛查任务中,提出了一种基于肺野分割的数据增强方法,该方法通过分别“擦除”左右肺部的随机区域增加了样本的多样性,从而提高了肺炎分类任务的准确率。实验表明,这种数据增强方法可以将新冠肺炎检出率至少提高2.2%。Because chest X-rays are affected by various tissue shadows and lesions,the results of current lung field segmentation algorithms often suffer from problems such as holes or unsmooth edges.To solve this problem,a lung field segmentation network based on multi-scale convolution and feature pyramid is proposed in this paper.This network uses multi-scale convolution module and multi-scale feature fusion module to extract and fuse multi-scale features.The Dice values of this network on JSRT and Montgomery datasets are 97.94%and 97.85%respectively and the pixel accuracy of this network on JSRT and Montgomery datasets are 98.76%and 98.96%respectively.Furthermore,we apply the lung field segmentation network to the pneumonia classification task,and propose a data augmentation method based on lung field segmentation.This data augmentation method increases the diversity of samples by erasing the random regions of the left and right lungs,thereby the accuracy of the pneumonia classification task is improved.Experiments show that this data augmentation method can increase the COVID-19 detection rates by at least 2.2%.

关 键 词：肺野分割 多尺度卷积 多尺度融合 肺炎筛查 数据增强 随机擦除 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]