基于胸部X线卷积神经网络模型诊断儿童不同病原体社区获得性肺炎

目的 评价基于胸部X线卷积神经网络（CNN）模型诊断儿童不同病原体社区获得性肺炎（CAP）的价值。方法 纳入1 769例CAP患儿，根据病原学诊断分为病毒组（n=487）、细菌组（n=496）及肺炎支原体（MP）组（n=786），对比组间胸部X线征象的差异；将患儿以7：1：2比例随机分为训练集、验证集和测试集，对测试集患儿根据性别和年龄分为不同亚组进行分层分析。基于胸部X线片分割全肺和病灶ROI，分别训练全肺模型和局部模型，通过混淆矩阵评估2种模型的整体效能；绘制受试者工作特征（ROC）曲线，计算曲线下面积（AUC），评价2种模型诊断不同病原CAP的效能；采用Delong检验比较模型诊断效能的差异。结果 3组病变累及范围、受累肺组织密度改变特点、肺过度通气及空洞差异均有统计学意义（P均<0.05）。全肺模型及局部模型诊断不同病原CAP的准确率分别为61.85%及58.04%，精确度分别为63.77%及54.05%。全肺模型和局部模型诊断MP性CAP的效能最佳，AUC分别为0.798及0.819；全肺模型诊断病毒及细菌性CAP的AUC均大于局部模型（P均<0.05）。全肺模型和局部模型诊断测试集中男性亚组和女性亚组不同病原CAP的AUC、诊断高年龄亚组和低年龄亚组病毒性及细菌性CAP的AUC差异均无统计学意义（P均>0.05），诊断高年龄亚组和低年龄亚组MP性CAP的AUC差异均有统计学意义（P均<0.05）。结论 基于胸部X线片建立CNN模型诊断儿童不同病原体CAP的效能较好；全肺模型优于局部模型，2种模型均对MP性CAP诊断效能最佳。

Chest X-ray model based on convolution neural network in diagnosis of children community acquired pneumonia caused by different pathogens

Objective To evaluate the value of chest X-ray model based on convolution neural network (CNN) in diagnosis of children community acquired pneumonia (CAP) caused by different pathogens. Methods A total of 1 769 CAP children were enrolled and divided into virus group (n=487),bacteria group (n=496) and mycoplasma pneumoniae (MP) group (n=786) according to the etiological diagnosis. The chest X-ray findings were compared among groups. Then the children were randomly divided into training set, verification set and test set at the ratio of 7:1:2. Children in the test set were further divided into subgroups according to gender and age for stratified analysis. Based on chest X-ray segmentation of whole lung and lesion''s ROI, the full lung model and local model were trained respectively, and the overall efficacy of these 2 models were evaluated with confusion matrix. Then the receiver operating characteristic (ROC) curves were drawn, and the areas under the curve (AUC) were calculated to evaluate the effectiveness of these 2 models in diagnosis of CAP caused by different pathogens. The Delong test was used to compare the diagnostic efficiency between the models.Results There were statistically differences of lesion involvement scope, characteristic of density change, hyperventilation and cavitation of the lungs among 3 groups (all P<0.05). The accuracy of the full lung model and the local model for diagnosis of CAP caused by different pathogens was 61.85% and 58.04%, and the local model was 63.77% and 54.05%, respectively. Both the full lung model and the local model had the best diagnostic efficacy for MP CAP, and the AUC was 0.798 and 0.819, respectively. The AUC of the full lung model in detecting viral and bacterial CAP were both higher than that of the local model (both P<0.05). There was no significant difference of AUC between models in diagnosing the pathogen of infection in the male subgroup and the female subgroup (all P>0.05), nor of AUC of 2 models for diagnosis of viral and bacterial CAP between the high age subgroup and the low age subgroup (all P>0.05), but there were significant differences of AUC of 2 models for diagnosing MP CAP (both P<0.05). Conclusion Chest X-ray CNN model had good diagnostic efficiency for CAP caused by different pathogens, and the full lung model was better than the local model. Both models had the best diagnostic efficiency for MP CAP.