基于多中心MRI的3D-ResNet101深度学习模型预测脑胶质瘤术前分级的研究  

The MRI-based 3D-ResNet101 deep learning model for predicting preoperative grading of gliomas:A multicenter study

作　　者：李大瑞 胡万均 刘光耀[1,2] 甘铁军 马来阳 张静 LI Darui;HU Wanjun;LIU Guangyao;GAN Tiejun;MA Laiyang;ZHANG Jing(Department of Magnetic Resonance,Lanzhou University Second Hospital,Lanzhou 730030,China;Gansu Province Clinical Research Center for Functional and Molecular Imaging,Lanzhou 730030,China;Second Clinical School,Lanzhou University,Lanzhou 730030,China)

机构地区：[1]兰州大学第二医院核磁共振科,兰州730030 [2]甘肃省功能及分子影像临床医学研究中心,兰州730030 [3]兰州大学第二临床医学院,兰州730030

出　　处：《磁共振成像》2023年第5期25-30,共6页Chinese Journal of Magnetic Resonance Imaging

基　　金：甘肃省卫生健康行业科研项目(编号:GSWSKY2020-68);兰州大学第二医院“萃英科技创新”项目(编号:CY2021-BJ-A05)。

摘　　要：目的 术前准确无创预测胶质瘤分级仍然具有挑战性。基于常规T2WI图像开发一种鲁棒性强的残差神经网络(Residual Networks,Res Net)深度学习模型以预测脑胶质瘤术前病理分级。材料与方法回顾性分析919例经病理证实为胶质瘤患者的术前T2WI图像,其中708例为2014年6月至2021年4月在兰州大学第二医院收治的患者数据,211例来源于癌症影像档案(The Cancer Imaging Archive,TCIA)数据库。TCIA数据集又被细分为开发集(n=135)和独立测试集(n=76),将兰州大学第二医院数据集和TCIA开发集的数据按7∶3随机分为训练集(n=590)和测试集(n=253),基于T2WI图像构建3 D-Res Net101深度学习模型。训练后的模型在测试集和独立测试集进行验证,并通过宏观F1分数、准确率(accaruy,ACC)及受试者工作特征(receiver operating characteristic,ROC)曲线对模型效能进行评估。结果基于T2WI构建的3 D-Res Net101深度学习模型在训练集和测试集ACC分别为99%、95%,F1分数分别为99%、95%,ROC曲线下面积(area under the curve,AUC)分别为0.98、0.97;独立测试集ACC为83%、F1分数为83%、AUC为0.89。结论 基于T2WI图像的3 D-Res Net101深度学习模型预测高、低级别胶质瘤具有较高的准确性、鲁棒性。该方法可用于术前胶质瘤分级的无创预测,并有助于提升患者临床管理的有效性。Objective:The preoperative accurate and non-invasive prediction of glioma grading remains challenging.Developing a robust Residual Networks(ResNet)deep learning model based on conventional T2WI images to predict preoperative pathological grading of gliomas.Materials and Methods:A retrospective analysis of preoperative T2WI images of 919 patients with pathologically confirmed glioma,of which 708 were data from patients enrolled at the Second Hospital of Lanzhou University from June 2014 to April 2021 and 211 were derived from The Cancer Imaging Archive(TCIA)database.The TCIA dataset was subdivided into a development set(n=135)and an independent test set(n=76).The data from the Second Hospital of Lanzhou University dataset and the TCIA development set were randomly split 7∶3 into a training set(n=590)and a test set(n=253)to construct a 3D-ResNet101 deep learning model based on T2WI images.After the training,the models were validated on the test set and independent test set,where the model efficacy was assessed by macro F1 scores,accuracy(ACC),and receiver operating characteristic(ROC)curves.Results:The 3D-ResNet101 deep learning model constructed based on T2WI had ACCs of 99%and 95%in the training and test sets,respectively;The F1 scores were 99%and 95%,respectively;the area under the ROC curve(AUC)were 0.98 and 0.97,respectively;the ACC of the independent test set was 83%,the F1 score was 83%,and the AUC was 0.89.Conclusions:The 3D-ResNet101 deep learning model based on T2WI images predicts high-and low-grade gliomas with high accuracy and robustness.The method can be used for the non-invasive prediction of preoperative glioma grading as well as helping to improve the effectiveness of clinical management of patients.

关 键 词：胶质瘤 3D-残差神经网络 深度学习 磁共振成像 T2加权成像 

分 类 号：R445.2[医药卫生—影像医学与核医学] R730.264[医药卫生—诊断学]