多标签病毒性肺炎图像分割方法的研究  

Research on Multi-Label Viral Pneumonia Image Segmentation Method

作　　者：吴怡红 杨勇[1] 叶宏伟 王小状 孙芳芳 Wu Yihong;Yang Yong;Ye Hongwei;Wang Xiaozhuang;Sun Fangfang(School of Automation,University of Hangzhou Dianzi University,Hangzhou 310018,China;Zhejiang Minfound Intelligent Healthcare Technology Co.,Ltd.,Hangzhou 310018,China)

机构地区：[1]杭州电子科技大学自动化学院,杭州310018 [2]浙江明峰智能医疗科技有限公司,杭州310018

出　　处：《中国生物医学工程学报》2023年第6期720-729,共10页Chinese Journal of Biomedical Engineering

摘　　要：目前,对于病毒性肺炎病灶分割的研究大多是基于单标签分割的,其分割准确率并不是特别理想。为提高病灶分割的准确率,提出了一种基于多标签的病毒性肺炎图像分割方法。本方法结合预先已知的肺部特征知识,将肺部特征信息叠加病灶特征信息形成多标签,并利用One-Hot编码原理和改进的3D-UNet网络模型,结合来源于大挑战比赛中的病毒性肺炎病灶分割挑战-2020公共数据集进行多标签图像分割训练,其中训练集139例,验证集和预测集同为40例,共计179例。在评价指标的选择上,除了常规的Dice评价指标外,本研究还提出了一个新的评估指标,炎肺比(focus-lung ratio),提供了病灶和肺部的比例,结合精度评价指标,可用于衡量模型的鲁棒性。最终,该研究提出的双标签分割方法在公开数据集上的预测集的Dice可达到70.10%,相较于同网络下的单标签分割方法,提高4.20%的准确率;该研究还通过3种不同网络(级联Res-UNet,nnUNet-ResUNet,3D-UNet)进行对比试验,实验结果表明在训练集上Dice损失达到最小,为5.00%;在验证集上Dice达到最大,为75.70%。所提出的方法改善了肺炎病灶分割的准确率,提高了模型的鲁棒性,对未来相关研究具有一定的临床价值和潜在意义。Researches on the segmentation algorithm of COVID-19’s lesion are mostly based on the singlelabel segmentation algorithm,but the accuracy can’t reach the clinical criteria.In this paper,a new method of COVID-19’s lesion segmentation based on multi-label was proposed,training on COVID-19 Lung CT Lesion Segmentation Challenge-2020 dataset in the Grand Challenge.The dataset contains 179 cases,including 139 cases in the training set and the rest 40 cases in both of the validation and prediction set.We conducted lung regions with existing lung region segmentation model,which generated from LUNA16 dataset.The generated lung region label was incorporated with the lesion label to form the multi-label of training dataset.The One-Hot coding principle and improved 3D-UNet network model is used for training.This paper also proposed a new evaluation index,focus-lung ratio which was used to reflect the proportion of lesion regions in the lung and measured the model’s robustness with other indicators.In the end,the prediction’s Dice reached 70.10%,which is 4.20% higher than the single-label segmentation method under the same network.Besides,our results were compared with some published data,and ours displayed better performance,the validation’s accuracy of dataset reached 75.70%.Experimental results showed that the proposed algorithm improved the accuracy of pneumonia lesion segmentation and the robustness of the model,therefore,is of clinical value and potential significance for future studies.

关 键 词：深度学习 图像分割 CT图像 多标签 病毒性肺炎 

分 类 号：R318[医药卫生—生物医学工程]