基于深度学习的脑部MRI解剖结构的语义分割  

Semantic image segmentation of brain MRI with deep learning

作　　者：胡益民[1] 赵惠萍[2] 李薇[1] 李军[1] HU Yimin;ZHAO Huiping;LIWei;LI Jun(Department of Neurology,Beijing Chuiyangliu Hosipital,Beijing 100022;Department of Oncology,Brain Hospital of Human Province,Changsha 410007,China)

机构地区：[1]北京市垂杨柳医院神经内科,北京100022 [2]湖南省脑科医院肿瘤科,长沙410007

出　　处：《中南大学学报（医学版）》2021年第8期858-864,共7页Journal of Central South University ：Medical Science

摘　　要：目的:既往脑部MRI图像分割法,如阈值法、边界检测法、区域法等,在复杂场景下分割效果不理想。本研究在深度学习分割技术的基础上,应用空洞卷积算法结合条件随机场(conditional random field,CRF)算法构造神经网络模型,对脑部MRI中的丘脑、尾状核和豆状核3种解剖结构进行分割,为脑部疾病MRI诊断打好基础。方法:随机选取1200张脑部MRI-Flair图像,人工标记出丘脑、尾状核和豆状核3种解剖结构,其中1000张作为训练数据集,200张作为测试数据集。采用深度卷积神经网络(deep convolutional neural networks,DCNN)结合CRF算法建立神经网络模型,将训练数据集输入模型,迭代30000次后得到参数化的神经网络模型。利用测试数据集评估、测试神经网络模型,并输出预测图像。结果:模型优化结果表明新构造的脑部MRI分割模型——DeepXAG的精确度最高,因此选用DeepXAG作为本研究的分割算法。DeepXAG模型的均交并比(mean intersection over union,mIOU)达到72.5%,明显高于其他经典分割算法(CRF-RNN、FCN-8s、DPN、RefineNet及PSPNet)。结论:DeepXAG模型分割脑部MRI的解剖结构具有较高的精确度及良好的鲁棒性。Objective:Previous studies on brain MRI image segmentation,such as threshold method,boundary detection method,and region method did not achieve good performance in complex scenes.Based on the deep learning segmentation technology,this study constructed a neural network model by using the algorithm of atrous convolution combined with conditional random field(CRF)to segment the thalamus,caudate nucleus,and lenticular nucleus in brain MRI,which laid a good foundation for MRI diagnosis of brain diseases.Methods:A total of 1200 MRI-Flair images of the brain were randomly selected,and 3 anatomical structures of thalamus,caudate nucleus,and lenticular nucleus were manually labeled,of which 1000 were used as training data sets and 200 were used as test data sets.The neural network model was established by using deep convolutional neural networks(DCNN)combined with CRF algorithm.The training data set was input into the model,and the parameterized neural network model was obtained after iteration for 30000 times.The test data set was used to evaluate,test,and output the predicted image.Results:The model optimization results showed that the new brain MRI segmentation model DeepXAG had the highest accuracy.Therefore,DeepXAG was selected as the segmentation algorithm.The mean intersection over union(mIOU)of the DeepXAG model was 72.3%,which was significantly higher than other classical segmentation algorithms(CRF-RNN1,FCN-8s2,DPN3,RefineNet4,and PSPNet5).Conclusion:The DeepXAG algorithm has good accuracy and robustness in segmenting the anatomical structure of brain MRI images.

关 键 词：图像语义分割 神经网络 空洞卷积 条件随机场 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程] TP391.41[自动化与计算机技术—控制科学与工程] R445.2[医药卫生—影像医学与核医学]