基于多模态MRI图像的3D卷积神经网络对肝纤维化分类的价值研究  

The value of 3D convolution neural network based on multimodal MRI images in the classification of liver fibrosis

作　　者：樊凤仙 胡万均 姜艳丽[1,2] 邹婕 杨品[1,2] 张静 FAN Fengxian;HU Wanjun;JIANG Yanli;ZOU Jie;YANG Pin;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)

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

出　　处：《磁共振成像》2022年第9期30-34,共5页Chinese Journal of Magnetic Resonance Imaging

基　　金：甘肃省科技计划项目(编号:21JR11RA122);兰州大学第二医院“萃英科技创新”计划(编号:CY2021-QN-B09)。

摘　　要：目的构建并验证多模态MRI图像3D卷积神经网络(convolutional neural network,CNN)模型对肝纤维化(liver fibrosis,LF)分类的价值。材料与方法回顾性分析经病理证实为LF,并行肝脏3.0 T MRI检查的224例LF患者的T1WI、T2WI、表观扩散系数(apparent diffusion coefficient,ADC)图像,按8∶2的比例随机分为训练集和测试集。对图像进行预处理后,应用训练集图像对模型进行网络结构迭代训练,建立3D-CNN深度学习模型对无显著LF(S0~S1)、显著LF(≥S2)进行分类。经过优化的CNN由三个卷积层、三个池化层和两个全连接层组成。训练完成后,用测试集数据对CNN模型进行测试,使用准确度(accuracy,ACC)曲线、损失函数(loss)曲线及受试者工作特征(receiver operating characteristic,ROC)曲线评价模型的性能。结果基于多模态MRI的3D-CNN深度学习模型在训练集中对LF分类的ROC曲线下面积(area under the curve,AUC)值为0.94,在测试集中的AUC为0.98。结论多模态3D-CNN深度学习模型可对无显著LF和显著LF进行分类,为LF的无创性评估提供更多选择。Objective:To construct a 3D convolution neural network(CNN)model of multi-modal MRI images,and verify its value in classification of liver fibrosis(LF).Materials and Methods:Two hundred and twenty four cases with LF confirmed by pathology were retrospectively collected.All patients underwent 3.0 T MRI exams.Collected the T1WI,T2WI,and apparent diffusion coefficient(ADC)images and randomly divided them into training group and testing group according to the ratio of 8∶2.After the images were preprocessed,the images of training group were used to iteratively train the network structure of the model.And then a 3D-CNN model was established to distinguish between no-significant LF(S0-S1)and significant LF(≥S2).The 3D-CNN model was composed of three convolution layers,three pooling layers and two fully connected layers.The accuracy(ACC),loss function curves and receiver operating characteristic(ROC)curves acquired by using the testing dataset were used to evaluate the performance of the 3D-CNN model.Results:The area under the curve(AUC)value of 3D-CNN model based on multiparametric MRI for LF classification was 0.94 in the training group and 0.98 in the testing group.Conclusions:The multiparametric 3D-CNN deep learning model may be an effective method,which can distinguish between no-significant and significant LF.It provides more options for non-invasive assessment of LF.

关 键 词：肝纤维化 多模态磁共振成像 机器学习 卷积神经网络 

分 类 号：R445.2[医药卫生—影像医学与核医学] R657.31[医药卫生—诊断学]