磁共振扩散加权成像纹理分析对新生儿低血糖脑损伤预后的预测价值

The texture analysis of MRI diffusion-weighted imaging for predicting prognosis of neonatal hypoglycemic encephalopathy

目的探讨磁共振扩散加权成像的纹理分析对新生儿低血糖脑损伤(HBD)预后的预测价值。方法回顾性分析2013年7月至2020年9月南京医科大学附属儿童医院收治的119例新生儿HBD患儿的临床与头磁共振成像检查资料。根据患儿7~8个月时贝利婴幼儿发展量表Ⅱ整体发育指数分判断,42例为神经系统正常(≥85分,A组)、46例为轻微发育迟缓(70~84分,B组)、31例为发育迟缓(≤69分,C组)。采用影像组学图像数据分析软件在扩散加权成像的表观扩散系数图上逐层勾画全脑区域(脑沟、池除外)作为感兴趣区,计算获得37个参数,并对3组患儿的临床资料、纹理参数(直方图特征、体积特征、灰度共生矩阵、灰度行程矩阵、邻域灰度差别矩阵、灰度区域矩阵)、临床参数联合纹理参数的诊断效能进行统计分析。采用多因素logistic回归分析差异有统计学意义的临床参数和纹理参数,并采用受试者工作特征(ROC)曲线评估其对新生儿HBD预后的预测效能。结果3组患儿的性别、胎龄、磁共振成像检查时年龄、分娩方式、血糖最低值差异均无统计学意义(均P>0.05)。A、B、C组出生体重[(3150±130)、(3020±220)和(2880±140)g,F=-3.10,P=0.015]、窒息史(10、18和20例,P=0.001)、妊娠期糖尿病/妊娠期高血压病史(14、29和21例,P=0.002)、低血糖持续时间[(5.0±0.2)、(8.0±0.4)和(14.0±1.7)d,F=-3.09,P=0.030]及新生儿神经行为评分[(32.0±3.2)、(28.0±2.6)和(22.0±1.9)分,F=-4.21,P=0.010],差异均有统计学意义。A、B、C组纹理参数的直方图峰度(2.57±1.12、3.66±0.98和4.23±0.37,F=3.54,P=0.010)和熵(5.89±1.09、7.67±2.12和8.92±1.62,F=-4.42,P=0.020),灰度共生矩阵的能量(0.48±0.01、0.36±0.02和0.23±0.01,F=-3.12,P=0.001),对比度(2419±21、3354±31和4313±26,F=-4.16,P=0.020)和异质性(126±14、153±23和344±43,F=-3.50,P<0.001);灰度行程矩阵中长行程优势(0.78±0.15、1.12±0.12和1.76±0.31,F=-4.13,P=0.006),长行程不均匀性(71.7±13.9、96.6±10.7和104.1±13.5,F=-0.98,P=0.001),行程百分比(0.91±0.05、0.84±0.21和0.72±0.17,F=2.97,P=0.010);邻域灰度差别矩阵的粗糙度(0.09±0.01、0.13±0.03和0.26±0.07,F=-1.95,P=0.003)和复杂度[0.16(0.04,4.14)、0.32(0.05,9.84)和0.45(0.15,10.14),H=-3.24,P=0.030],灰度区域矩阵的短区域因子(4.74±0.45、3.44±1.03和1.88±0.67,F=-3.14,P=0.040)和短区域高灰度级优势(278963±239、164607±544和111653±618,F=-3.84,P=0.001),差异均有统计学意义。logistic回归分析显示,低血糖持续时间、新生儿神经行为评分、能量、峰度、行程百分比和短区域因子是新生儿HBD预后不良的独立相关因素(OR值分别为7.43,4.09,1.10,2.11,1.36和1.68,P值分别为0.002,0.027,0.001,0.006,0.007和0.010)。ROC曲线显示将低血糖持续时间、新生儿神经行为评分与纹理参数联合,曲线下面积最大(AUC=0.94,P<0.001),诊断价值最大。结论磁共振扩散加权成像的纹理分析可早期预测HBD的预后,且联合临床特征预测效能更优化。

Objective To investigate the prognostic value of texture analysis of MRI diffusion weighted imaging(DWI)for neonatal hypoglycemic encephalopathy(HE).Methods The clinical data and MRI data of 119 patients with neonatal HE admitted to Children′s Hospital of Nanjing Medical University from July 2013 to September 2020 were retrospectively analyzed.The children were followed up to 7—8 months and scored by Bayley scales of infant and toddler development.According to the overall development index,the children were divided into three groups:normal group(≥85,group A,n=42),mild developmental retardation group(70-84,group B,n=46)and developmental retardation group(≤69,group C,n=31).The whole brain region(except sulcus and cisterna)was delineated as region of interest(ROI)by LIFEx 3.4 software in MRI apparent diffusion coefficient images.A total of 37 parameters were calculated automatically by the software,The clinical data,including gender,gestational age,age at MRI scan,birth weight,mode of delivery,history of asphyxia at birth,maternal preeclampsia or diabetes,minimum blood glucose,duration of hypoglycemia,neonatal behavioral neurological assessment(NBNA),presence or absence of polycythemia);the texture parameters,including histogram,volume,gray level co-occurrence matrix(GLCM),gray level run length matrix(GLRLM),neighborhood gray tone difference matrix(NGTDM),gray level size zone matrix(GLSZM),in the three groups were analyzed;and the diagnostic efficacy of clinical parameters and texture parameters was analyzed.Multivariate Logistic regression was used to analyze statistically significant clinical parameters and texture parameters,and receiver operating characteristic curve(ROC)was used to evaluate the prognostic efficacy of these parameter for neonatal HE.Results There were no significant differences in gender,gestational age,age at MRI scan,delivery mode and blood glucose minimum among the three groups(P>0.05).There were significant differences in birth weight[(3150±130)g,(3020±220)g,(2880±140)g,F=-0.31,P=0.015],history of suffocation(10 cases,18 cases,20 cases,P=0.001),history of maternal diabetes or preeclampsia(14 cases,29 cases,21 cases,P=0.002),blood glucose duration[(5.0±0.2)d,(8.0±0.4)d,(14.0±1.7)d,F=-3.09,P=0.030]and NBNA scores(32.0±3.2,28.0±2.6,22.0±1.9,F=-4.21,P=0.010)among three groups.There were significant differences in kurtosis and entropy of histogram(2.57±1.12,3.66±0.98,4.23±0.37,F=3.54,P=0.010;5.89±1.09,7.67±2.12,8.92±1.62,F=-4.42,P=0.020);energy,contrast and dissimilarity of GLCM(0.48±0.01,0.36±0.02,0.23±0.01,F=-3.12,P=0.001;2419±21,3354±31,4313±26,F=-4.16,P=0.020;126±14,153±23,344±43,F=-3.50,P<0.001);long run emphasis of GLRLM(0.78±0.15,1.12±0.12,1.76±0.31,F=-4.13,P=0.006),run length non-uniformity and run percentage(71.7±13.9,96.6±10.7,104.1±13.5,F=-0.98,P=0.001;0.91±0.05,0.84±0.21,0.72±0.17,F=2.97,P=0.010);coarseness and busyness of NGTDM[0.09±0.01,0.13±0.03,0.26±0.07,F=-1.95,P=0.003;0.16(0.04,4.14),0.32(0.05,9.84),0.45(0.15,10.14),H=-3.24,P=0.030],short-zone emphasis and short-zone high gray length emphasis of GLSZM(4.74±0.45,3.44±1.03,1.88±0.67,F=-3.14,P=0.040;278963±239,164607±544,111653±618,F=-3.84,P=0.001)among three groups.Multivariate Logistic regression showed that duration of hypoglycemia,NBNA score,energy,kurtosis,run percentage and short zone effect were independent risk factors for poor prognosis of neonatal HE(OR=7.43,4.09,1.10,2.11,1.36,1.68,P=0.002,0.027,0.001,0.006,0.007,0.010,respectively).ROC curve showed that for combined hypoglycemic duration,NBNA and texture parameters,the area under the curve(AUC)was the highest(AUC=0.94,P<0.001).Conclusion Texture analysis of the MRI diffusion weighted imaging can predict the prognosis of neonatal hypoglycemic encephalopathy at an early stage,which has better prediction efficiency when combined with clinical features.