基于支持向量机的MRI灌注与扩散技术对脑胶质瘤术前分级预测模型研究

Research on the Predictive Model of Preoperative Grading of Glioma by MRI Perfusion and Diffusion Technology: Based on Support Vector Machine

目的采用基于支持向量机(SVM)的方法,对动态磁敏感对比增强灌注加权成像(DSC-PWI)、3D动脉自旋标记(3D-ASL)和多b值扩散加权成像(DWI)三种功能磁共振成像(fMRI)技术进行建模分析,以期构建一种效能最高的胶质瘤术前分级预测模型,为临床医师对胶质瘤患者术前分级的评估和治疗方案的确定提供更多有用的信息和帮助。方法回顾性分析50例被病理证实胶质瘤患者的影像学检查资料。测量患者DSC-PWI得到瘤侧和对应健康侧的脑血容量(CBV)和脑血流量(CBF)值,3D-ASL得到瘤侧和对应健康侧的CBF值。多b值DWI得到瘤侧和对应健康侧的ADC_(stand) 、ADC_(fast)、ADC_(slow)、灌注分数(f)和水通道蛋白(AQP)值。采用特征递归消除(RFE)算法对变量进行筛选后建立SVM预测模型并与常规预测模型比较,并用受试者工作特征曲线(ROC)曲线下面积(AUC)评价模型预测性能以及采用决策曲线(DCA)分析其临床实用价值。结果年龄、性别以及MRI灌注参数ASL-CBF、ASL-CBF_(健侧)、DSC-CBF、DSC-CBV、ADC_(stand)和f_(健侧)在高低级别胶质瘤中分布差异有统计学意义(P<0.05)。根据RFE算法发现最终纳入模型的8个变量分别为:rASL-CBF、DSC-CBV、DSC-CBF、ADC_(stand)、Age、rf、AQP值和ADC_(slow-bi)。ROC曲线分析发现SVM模型预测效能和精度最高,其次是基于RFE法建立的Logistic模型,而采用患者瘤体侧与健康侧的相对值等传统的方法建立的DSC、DWI和3D-ASL常规预测模型其预测效能和精确度都较低;其AUC分别为0.969(95%CI:0.931~1.000)、0.879(95%CI:0.786~0.969)、0.767(95%CI:0.630~0.903)、0.696(95%CI:0.542~0.849)和0.836(95%CI:0.724~0.947)。PCA分析表明,概率阈值从5%起采用基于SVM的术前预测模型可以使得患者净获益,且当概率阈值≥10%时,SVM的临床获益显著要优于基于Logistic回归建立的常规预测模型。结论 DSC-PWI、3D-ASL和多b值DWI三种fMRI技术对高低级别的鉴别诊断均具有一定的预测价值,而采用SVM构建的预测模型与常规预测模型比较具有更好的预测能力。利用SVM预测模型的指标重要性分布图的可视化特点,可清晰地了解那些参数在模型的重要程度,有助于临床影像科医师对胶质瘤的级别加以判断。

Objective Using SVM method to model and analyze the three fMRI techniques of DSC,3 D-ASL and multi-b value DWI,constructing a most effective for clinicians about tumor grading in gliomas before surgery, which will provide more useful information and help. Methods 50 pathologically confirmed glioma patients who underwent 3 D-ASL,DSC-PWI,and multi-b value on MR scanner were retrospectively reviewed.DSC-PWI obtained the CBV and CBF values of the tumor side and the corresponding healthy side, and 3 D-ASL obtained the CBF value of the tumor side and the corresponding healthy side.Multi-b value DWI obtained ADC_(stand),ADC_(fast),ADC_(slow),F and AQP values of the tumor side and the corresponding healthy side.After screening the variables with the RFE algorithm, the SVM prediction model was established and compared with the normalized models, and The prediction performance of the model was evaluated by the area under the ROC Curve, and the DCA was used to analyze its clinical practical value. Results There were statistically significant differences in age, gender, and MRI perfusion parameters ASL-CBF,ASL-CBF healthy side, DSC-CBF,DSC-CBV,ADC_(stand),and F healthy side in BG of high and low grades(all P<0.05).The 8 variables of the optimal model found by the RFE algorithm were: rASL-CBF,DSC-CBV,DSC-CBF,ADC_(stand),Age, rF, AQP,and ADC_(slow)-bi.The ROC showed that the SVM model had the highest predictive power and accuracy, followed by the Logistic model based on the RFE algorithm, and the DSC,DWI and 3 D-ASL conventional predictions established by traditional Logistic methods such as the relative value of the patient’s tumor and the contralateral healthy side, which had lower predictive power and accuracy.The AUC were 0.969(95%CI:0.931~1.000),0.879(95%CI:0.786~0.969),0.767(95%CI:0.630~0.903),0.696(95%CI:0.542~0.849),and 0.836(95%CI:0.724~0.947),respectively.The DCA showed that the preoperative prediction model based on SVM with probability thresholds from 5% could bring net benefits to patients, and when the probability threshold ≥10%,the clinical benefits of SVM was significantly better than the conventional prediction model based on Logistic regression. Conclusion Three fMRI techniques of DSC-PWI,3 D-ASL and multi-b-value DWI have certain predictive value for grading in gliomas, and the predictive model constructed by SVM has better predictive ability compared with the normal models.By using the visual features of the index importance distribution map of the SVM prediction model, we can clearly understand the importance of those parameters in the model, which is helpful for clinical imaging doctors to judge the grade of glioma.