摘要
目的:探讨基于VGG19 的深度迁移学习模型在儿童胸部 X 线片(胸片)早期诊断肺炎支原体肺炎(MPP)中的价值.方法:搜集常州市第二人民医院儿科就诊的患儿 3763 例,其临床及影像资料齐全.将患儿胸片分成细菌性肺炎组、MPP组和病毒性肺炎组,并按照 9:1 随机分为训练集和测试集.另外选择 150 例患儿胸片作为验证集(细菌性肺炎、MPP 和病毒性肺炎各 50 例).模型性能评价指标包括在训练集和测试集中的准确率(Ac)及在验证集中的精确度(Pr)、召回率(Rc)、F1 评分(F1)和ROC曲线下面积(AUC).结果:VGG19 在训练集和测试集中的最高 Ac分别为 0.99 和 0.95.细菌性肺炎组的Pr、Rc、F1、AUC分别为 0.87、0.82、0.85、0.92,MPP组分别为 0.85、0.74、0.78、0.90,病毒性肺炎组分别为 0.76、0.88、0.81、0.93.VGG19 对三组图像分类诊断的总体 AUC 为 0.92.结论:基于VGG19 的深度迁移学习模型是儿童胸部X线片诊断MPP的可靠方法,能够帮助临床早期诊断MPP.
Abstract
Objective:This study was aimed to investigate the diagnostic value of a deep transfer learning method based on chest X-ray images for mycoplasma pneumonia(MPP)in children.Methods:A total of 3763 cases from the Department of Pediatrics of the Second People's Hospital of Changzhou were recruited in this study.Clinical data and radiographic data of each patient were completely recor-ded.The chest X-ray images were divided into bacterial,MPP,and viral groups.Each group was ran-domly divided into a train set and test set by a ratio of 9:1.Deep transfer learning based on the VGG19 model was established.Also,another 150 chest X-ray image samples with bacterial pneumonia,MPP and viral pneumonia were selected as a validation set:50 cases each.Metrics for model performance e-valuation include accuracy in the train set and test set,precision(Pr),recall rate(Rc),F1 score(F1)and area under the ROC curve(AUC)in the validation set.Results:Accuracies of VGG19 on the train set and the test set were 0.99 and 0.95,respectively.The values of Pr,Rc,F1 and AUC were bacterial group(0.87,0.82,0.85,0.92),MPP group(0.85,0.74,0.78,0.90)and viral group(0.76,0.88,0.81,0.93),respectively.The VGG19 model achieved an overall AUC of 0.92 for three groups.Conclusion:The VGG19 model is a reliable method to classify chest X-ray images of MPP children,which is a promising AI-assisted diagnosis method for early diagnosis of MPP.
基金项目
常州市科技局应用基础研究项目(CJ20220260)
NETL
NSTL
维普
万方数据