基于一般线性模型的机器学习方法在BOLD-fMRI脑胶质瘤患者个体化运动功能区定位中的应用价值

目的 探讨基于一般线性模型（GLM）的机器学习方法在血氧水平依赖功能磁共振成像（BOLD-fMRI）定位脑胶质瘤患者个体化运动功能中的应用价值。方法 前瞻性研究。纳入2017年11月—2021年11月西安交通大学第一附属医院神经外科确诊为脑胶质瘤且病灶位于大脑运动功能区的38例患者作为机器学习模型的验证集（男25例、女13例,年龄24~69岁）,同期招募健康志愿者50例作为模型的训练集（男26例、女24例,年龄22~68岁）。采用独立成分分析法（ICA）,随机提取98例人类连接组计划（HCP）受试者的静息态功能核磁共振（rs-fMRI）特征。依据健康志愿者的rs-fMRI和基于任务的功能磁共振（tb-fMRI）的相关性,训练基于GLM的机器学习模型。观察项目：（1）采用Pearson相关系数（CC）分析比较GLM预测的激活与实际激活的相似度。（2）采用Dice系数（DC）作为模型预测效能的定量指标,比较GLM与ICA方法的预测效能。结果 （1）胶质瘤患者基于GLM的机器学习方法所预测的激活与实际tb-fMRI的功能激活相似度高（89.47% （34/38）的患者CC值>0.30）]。（2）胶质瘤患者GLM预测任务态运动功能激活的效能,DC为0.34（0.27,0.42）,优于ICA方法的效能DC 0.26（0.16,0.30）,差异有统计学意义（Z=-3.88,P<0.001）;GLM在肿瘤半球的预测效能优于ICA方法,DC分别为0.36（0.17,0.48）和0.34（0.04,0.45）,差异有统计学意义（Z=-2.43,P=0.015）;2种方法在非肿瘤半球的预测效果均显著高于肿瘤半球（Z=-4.33、-3.59,P值均<0.001）。结论 基于GLM的机器学习方法能够很好地在术前利用rs-fMRI数据预测出胶质瘤患者的tb-fMRI运动功能激活,且其预测效果好于ICA方法。

Application value of a machine learning method based on general linear model in the localization of individual motor function in patient with glioma after blood oxygen level dependent functional magnetic resonance imaging

Objective The study aimed to evaluate the application value of a machine learning method based on general linear model (GLM) in the localization of individual motor function in patients with glioma after blood oxygen level dependent functional magnetic resonance imaging (BOLD-fMRI). Methods A retrospective study was conducted, and strict clinical screening was performed in the Neurosurgery Department of the First Affiliated Hospital of Xi'an Jiaotong University from November 2017 to November 2021. A total of 38 pathologically confirmed patients with glioma located in the motor area were selected and included in the validation set of the machine learning model (25 males, 13 females; aged 24-69), and 50 healthy volunteers were recruited and included in the training set (26 males, 14 females; aged 22-68). Extracting the resting-state fMRI (rs-fMRI) features from 98 subjects in the Human Connectome Project (HCP) using the independent component analysis (ICA). A machine learning model based on GLM was trained using the correlation between the rs-fMRI and task-based fMRI (tb-fMRI) features of healthy subjects. (1) GLM-predicted activation and actual activation were compared by Pearson correlation coefficient (CC) analysis; (2) the dice coefficient (DC) was used as a quantitative indicator of the prediction efficiency of the model and used in comparing the prediction efficiency of GLM and ICA methods. Results (1) GLM-prediction activation in glioma patients was highly similar to task-state function activation (CC>0.30 in 89.47% 34/38] of patients). (2) GLM was better than ICA in predicting task-state motor function activation. The DC was 0.34 (0.27, 0.42), and 0.26 (0.16, 0.30), respectively, the difference was statistically significant (Z=-3.88; P<0.001). In the tumor-containing hemisphere, GLM was better than ICA in predicting task-state activation, with DCs of 0.36 (0.17, 0.48) and 0.34 (0.04, 0.45), respectively (Z=-2.43, P=0.015). The prediction effects of the two methods in the nontumor hemisphere was significantly higher than that in the tumor hemisphere (Z=-4.33, -3.59; all P values<0.001). Conclusion GLM-based machine learning can predict tb-fMRI motor activation in patients with glioma after rs-fMRI and before surgery and is more efficient than ICA.