临床因素模型、基于基线平扫CT的深度学习模型及两者联合预测脑出血血肿扩大的价值

The value of clinical model,deep learning model based on baseline noncontrast CT and the combination of the two in predicting hematoma expansion in cerebral hemorrhage

目的探讨临床因素模型、基于基线平扫CT的深度学习模型及两者联合预测脑出血血肿扩大的价值。方法该研究为横断面研究。回顾性分析2017年1月至2021年12月在苏州大学附属第二医院首次就诊的脑出血患者471例,采用随机函数以7∶3的比例分为训练集(330例)和验证集(141例)。所有患者均在24 h内接受2次CT检查且以血肿体积增加>33%或血肿量绝对增加>6 ml为血肿扩大。按照有无出现血肿扩大将所有患者分为血肿扩大组与血肿无扩大组,采用两样本t检验、Mann-Whitney U检验或χ2检验进行单因素分析,将差异有统计学意义的指标纳入多因素logistic回归分析,筛选出与血肿扩大有关的独立影响因素并建立临床因素模型。应用ITK-SNAP软件在平扫CT图像上对脑出血灶进行手动标记、分割,训练并建立基于ResNet50架构的深度学习模型。将独立临床影响因素与基于深度学习模型计算得出的深度学习评分联合建立预测脑出血血肿扩大的联合模型。在训练集和验证集中采用受试者操作特征(ROC)曲线和决策曲线评价临床因素模型、深度学习模型及联合模型预测脑出血血肿扩大的价值。结果471例脑出血患者中血肿扩大组136例、血肿无扩大组335例。多因素logistic分析显示男性(OR=1.790,95%CI 1.136~2.819,P=0.012)、发病时间(OR=0.812,95%CI 0.702~0.939,P=0.005)、口服抗凝药物史(OR=2.157,95%CI 1.100~4.229,P=0.025)、入院格拉斯哥昏迷评分(OR=0.866,95%CI 0.807~0.929,P<0.001)、红细胞分布宽度(OR=1.045,95%CI 1.010~1.081,P=0.011)是预测脑出血血肿扩大的独立影响因素。ROC曲线分析显示,训练集中临床因素模型、深度学习模型及联合模型预测脑出血血肿扩大的曲线下面积(AUC)分别为0.688(95%CI 0.635~0.738)、0.695(95%CI 0.642~0.744)和0.747(95%CI 0.697~0.793),其中联合模型AUC优于临床因素模型(Z=2.54,P=0.011)、深度学习模型(Z=2.44,P=0.015)。在验证集中,临床因素模型、深度学习模型及联合模型预测脑出血血肿扩大的AUC分别为0.687(95%CI 0.604~0.763)、0.683(95%CI 0.599~0.759)和0.736(95%CI 0.655~0.806),两两比较差异均无统计学意义。决策曲线分析显示联合模型的净获益率最高,其临床实用性较强。结论该研究建立的深度学习模型与临床因素模型对脑出血的血肿扩大均有一定的预测价值;二者联合建立的联合模型预测价值最高,可应用于预测血肿扩大。

ObjectiveTo investigate the predictive value of clinical factor model,deep learning model based on baseline plain CT images,and combination of both for predicting hematoma expansion in cerebral hemorrhage.MethodsThe study was cross-sectional.Totally 471 cerebral hemorrhage patients who were firstly diagnosed in the Second Affiliated Hospital of Soochow University from January 2017 to December 2021 were collected retrospectively.These patients were randomly divided into a training dataset(n=330)and a validation dataset(n=141)at a ratio of 7∶3 by using the random function.All patients underwent two noncontrast CT examinations within 24 h and an increase in hematoma volume of>33%or an absolute increase in hematoma volume of>6 ml was considered hematoma enlargement.According to the presence or absence of hematoma enlargement,all patients were divided into hematoma enlargement group and hematoma non-enlargement group.Two-sample t test,Mann-Whitney U test or χ^(2) test were used for univariate analysis.The factors with statistically significant differences were included in multivariate logistic regression analysis,and independent influences related to hematoma enlargement were screened out to establish a clinical factor model.ITK-SNAP software was applied to manually label and segment the cerebral hemorrhage lesions on plain CT images to train and build a deep learning model based on ResNet50 architecture.A combination model for predicting hematoma expansion in cerebral hemorrhage was established by combining independent clinical influences with deep learning scores.The value of the clinical factor model,the deep learning model,and the combination model for predicting hematoma expansion in cerebral hemorrhage was evaluated using receiver operating characteristic(ROC)curves and decision curves in the training and validation datasets.ResultsAmong 471 cerebral hemorrhage patients,136 cases were in the hematoma enlargement group and 335 cases were in the hematoma non-enlargement group.Regression analyses showed that male(OR=1.790,95%CI 1.136-2.819,P=0.012),time of occurrence(OR=0.812,95%CI 0.702-0.939,P=0.005),history of oral anticoagulants(OR=2.157,95%CI 1.100-4.229,P=0.025),admission Glasgow Coma Scale score(OR=0.866,95%CI 0.807-0.929,P<0.001)and red blood cell distribution width(OR=1.045,95%CI 1.010-1.081,P=0.011)were the independent factors for predicting hematoma expansion in cerebral hemorrhage.ROC curve analysis showed that in the training dataset,the area under the curve(AUC)of clinical factor model,deep learning model and combination model were 0.688(95%CI 0.635-0.738),0.695(95%CI 0.642-0.744)and 0.747(95%CI 0.697-0.793)respectively.The AUC of the combination model was better than that of the clinical model(Z=0.54,P=0.011)and the deep learning model(Z=2.44,P=0.015).In the validation dataset,the AUC of clinical factor model,deep learning model and combination model were 0.687(95%CI 0.604-0.763),0.683(95%CI 0.599-0.759)and 0.736(95%CI 0.655-0.806)respectively,with no statistical significance.Decision curves showed that the combination model had the highest net benefit rate and strong clinical practicability.ConclusionsBoth the deep learning model and the clinical factor model established in this study have some predictive value for hematoma expansion in cerebral hemorrhage;the combination model established by the two together has the highest predictive value and can be applied to predict hematoma expansion.