T1WI深度学习模型评估新生儿高胆红素血症脑损伤  

T1WI deep learning models for evaluating brain injury of neonatal hyperbilirubinemia

作　　者：崔静威 牛永超[2] 谢北辰 刘畅 段金辉 薛芹 闫瑞芳[1] CUI Jingwei;NIU Yongchao;XIE Beichen;LIU Chang;DUAN Jinhui;XUE Qin;YAN Ruifang(Department of Magnetic Resonance,the First Affiliated Hospital of Xinxiang Medical University,Xinxiang 453100,China;Department of Magnetic Resonance,Xinxiang Central Hospital,Xinxiang 453000,China)

机构地区：[1]新乡医学院第一附属医院磁共振科,河南新乡453100 [2]新乡市中心医院磁共振科,河南新乡453000

出　　处：《中国医学影像技术》2025年第3期394-398,共5页Chinese Journal of Medical Imaging Technology

基　　金：2023年度河南省医学科技攻关计划联合共建项目(LHGJ20230505);2023年度新乡市市级重点实验室(新乡市新生儿神经系统疾病影像研究重点实验室)。

摘　　要：目的观察T1WI深度学习模型评估新生儿高胆红素血症(NHB)脑损伤的价值。方法收集中心A 106例NHB患儿(新生儿行为神经测定评分≤37,NHB组)及119名非NHB新生儿(对照组),以及中心B 34例NHB患儿及18名非NHB新生儿;于颅脑T1WI中沿双侧苍白球勾画ROI,对中心A数据进行预处理获得690片切片,按8∶2比例分为训练集(n=552)与测试集(n=138),分别建立ResNet18、DenseNet121及EfficientNetB0模型;以中心B数据为验证集进行外部验证。绘制受试者工作特征曲线,计算曲线下面积(AUC),与传统目视分析方法比较,评价模型评估NHB脑损伤的效能。结果ResNet18模型评估NHB脑损伤的AUC为0.910~0.990,显著高于DenseNet121模型(0.710~0.820)及EfficientNetB0模型(0.640~0.740)(P均<0.001)。ResNet18模型评估NHB脑损伤的准确率、敏感度及精确度均高于目视分析(P均<0.05),二者特异度差异无统计学意义(P>0.05)。结论T1WI ResNet18模型用于评估NHB脑损伤效能优异,且泛化能力良好。Objective To observe the value of T1WI deep learning models for evaluating brain injury of neonatal hyperbilirubinemia(NHB).Methods Totally 106 NHB(defined as newborns with neonatal behavioral neurological assessment≤37,NHB group)and 119 non-NHB newborns(control group)in center A,as well as 34 NHB and 18 non-NHB newborns in center B were collected.ROI was delineated based on bilateral globus pallidus on T1WI.A total of 690 slices were obtained by preprocessing data of center A and then were divided into training set(n=552)and test set(n=138)at a ratio of 8∶2.ResNet18,DenseNet121 and EfficientNetB0 models was established,respectively.External validation was performed based on data of center B.Receiver operating characteristic curves were drawn,area under the curves(AUC)were calculated to evaluate the performance of models for assessing NHB brain injuries compared with traditional visual analysis.Results The AUC of ResNet18 model for evaluating NHB brain injury was 0.910—0.990,significantly higher than that of DenseNet121 model(0.710—0.820)and EfficientNetB0 model(0.640—0.740)(all P<0.001).The accuracy,sensitivity and precision of ResNet18 model for evaluating NHB brain injury were all higher than those of visual analysis(all P<0.05),while no significant difference of specificity was found between the above two(P>0.05).Conclusion T1WI ResNet18 model showed excellent performance and generalization ability for evaluating NHB brain injury.

关 键 词：高胆红素血症 新生儿 脑损伤 磁共振成像 深度学习 

分 类 号：R722.17[医药卫生—儿科] R445.2[医药卫生—临床医学]