机器学习结合MRI影像鉴别脑转移瘤来源  

Machine learning integrating MRI for distinguishing the origin of brain metastases

作　　者：崔婀娜[1] 王晓煜[2] 赵鹏[3] 姜明辰 CUI Enuo;WANG Xiaoyu;ZHAO Peng;JIANG Mingchen(School of Intelligent Science and Engineering,Shenyang University,Shenyang 110044,Liaoning Province,China;Department of Radiology,Cancer Hospital of China Medical University,Shenyang 110042,Liaoning Province,China;Department of Medical Imaging,Cancer Hospital of China Medical University,Shenyang 110042,Liaoning Province,China;School of Intelligent Medicine,China Medical University,Shenyang 110122,Liaoning Province,China)

机构地区：[1]沈阳大学智能科学与工程学院,辽宁沈阳110044 [2]辽宁省肿瘤医院(中国医科大学肿瘤医院)放射科,辽宁沈阳110042 [3]辽宁省肿瘤医院(中国医科大学肿瘤医院)医学影像科,辽宁沈阳110042 [4]中国医科大学智能医学学院,辽宁沈阳110122

出　　处：《肿瘤影像学》2024年第6期569-576,共8页Oncoradiology

基　　金：国家重点研发项目:BTIT(2022YFF1202803);辽宁省教育厅面上项目(JYTMS20230132)。

摘　　要：目的:通过脑转移瘤患者的磁共振成像(magnetic resonance imaging,MRI)图像建立机器学习模型,鉴别脑转移瘤(brain metastases,BM)的原发病灶来源。方法:本研究数据取自2017年1月—2020年9月就诊于辽宁省肿瘤医院的肿瘤脑转移患者。由经验丰富的影像科医师对患者的脑转移瘤活性区进行手动勾画,通过计算机方法得到包含活性区和瘤周区域的4 mm环形区域,从对比增强T1加权成像(contrast-enhanced T1-weighted imaging,CE-T1WI)和T2加权成像(T2-weighted imaging,T2WI)序列中提取影像组学特征,合并序列后经过U检验、最小绝对收缩和选择算子(least absolute shrinkage and selection operator,LASSO)算法、赤池信息量准则(Akaike information criterion,AIC)进行三步法特征筛选,并绘制受试者工作特征(receive operating characteristic,ROC)曲线,计算曲线下面积(area under curve,AUC)作为鉴别模型分类性能的指标。结果:共纳入215例脑转移瘤患者,包括肺来源100例,乳腺来源50例,肠胃来源50例,其他来源15例。从两个序列中分别提取1967个影像组学特征,成功构建3个影像组学模型,分别是鉴别脑转移瘤是否来自于肺癌病灶的RS_Lung模型,是否来自乳腺癌病灶的RS_Breast模型和是否来自胃肠癌病灶的RS_Gastrointestic模型,其训练集AUC分别为0.898、0.872和0.938,灵敏度分别为0.908、0.744和0.860,特异度分别为0.818、0.879和0.909。结论:本研究基于4 mm环形区域构建的影像组学模型在脑转移瘤来源的鉴别任务中取得了良好结果,有可能成为无创的术前新标志物,指导脑转移瘤患者的个性化治疗方案。Objective:Based on magnetic resonance imaging of patients with brain metastases,using imaging omics methods to differentiate the primary lesion of brain metastases and determine the specific source of the patient’s brain metastases.Methods:The data of this study were collected from patients with brain metastasis who treated in Liaoning Cancer Hospital from January 2017 to September 2020.Experienced imaging doctors manually delineate the active area of the patient’s brain metastasis tumor,and obtain a 4 mm circular area containing the active area and the surrounding area through computer methods.And 1967 imaging omics features were extracted from each region of the contrast-enhanced T1-weighted imaging(CE-T1WI)and T2-weighted imaging(T2WI)sequences.After merging the sequences,a three-step feature screening method was performed using U-test,least absolute shrinkage and selection operator(LASSO)logistic regression,and Akaike information criterion(AIC),draw the receive operating characteristic(ROC)curve to calculate the area under curve(AUC)value as an indicator of the classification performance of the discriminative model.Results:A total of 215 patients with brain metastases were identified,including 100 cases of lung origin,50 cases of breast origin,50 cases of gastrointestinal origin,and 15 cases of other origin.From each of the two sequences,1967 imaging histological features were extracted. Three radiomics models were successfully constructed to differentiate whether brain metastases arose fromlung cancer lesions in the RS_Lung model, breast cancer lesions in the RS_breast model and gastrointestinal cancer lesions in theRS_Gastrointestic model, with training set AUC of 0.898, 0.872 and 0.938, sensitivity of 0.908, 0.744 and 0.860, specificity of 0.818,0.879 and 0.909, respectively. Conclusion: This study achieved good results in the distinguishing task derived from brain metastasesusing an imaging omics model constructed based on a 4 mm annular region, which has the potential to be a noninvasive preoperative

关 键 词：脑转移瘤 磁共振成像 机器学习 

分 类 号：R739.41[医药卫生—肿瘤] R445.2[医药卫生—临床医学]