基于深度学习的肺炎病灶分割技术用于新型冠状病毒肺炎的定量分析

Deep Learning Based Pneumonia Lesion Segmentation Used for Quantitative Analysis of Coronavirus Disease 2019

目的:探究基于深度学习的肺炎病灶分割模型用于新型冠状病毒肺炎(COVID-19)CT影像定量分析的可行性。方法:基于深度学习技术,以73例肺炎CT影像和分割标记训练病灶分割模型,13例肺炎CT影像测试模型,217例肺炎/非肺炎CT影像训练分类模型,56例CT影像测试分类模型。用肺炎病灶分割模型在28例COVID-19疑似病例(7例阳性,21例阴性)CT影像中分割病灶区域,结合肺叶分割结果汇总定量。结果:肺炎病灶分割模型在同源测试集上精准度为75.4%,对COVID-19病灶的分割精准度为67.7%。C0VID-19阳性病例相较于阴性病例,病灶密度均值更高(P=0.04),且受累肺叶数量更多(P=0.01)。结论:基于深度学习的肺炎病灶分割模型能用于COVID-19病灶分割,分割后定量分析有利于明确COVID-19的CT影像特征。

Purpose:To explore the feasibility of deep learning based pneumonia lesion segmentation model used for quantitative analysis of coronavirus disease 2019(COVID-2019)in CT images.Methods:Based on deep learning,the lesion segmentation model was trained with 73 CT images of pneumonia and segmentation annotations,and was tested with 13 CT images cases of pneumonia;the classification model was trained with 217 CT images and tested with 56 CT images.The lesions were outlined using the pneumonia lesion segmentation model in 28 CT images of suspected COVID-19(7 RT-PCT positive cases,21 negative),and the results were quantified with pulmonary lobe segmentation.Results:The precision of the pneumonia lesion segmentation model on testing data was 75.4%,and the precision for the lesions of COVID-19 was 67.7%.Compared with negative cases,the positive cases of COVID-19 have a higher mean density of lesions(P=0.04)and a greater number of affected lung lobes(P=0.01).Conclusions:The deep learning based pneumonia lesion segmentation model can be used for lesions segmentation of COVID-19.and the quantitative analysis after segmentation are helpful for defining the CT imaging characteristics of COVID-19.