基于深度学习MR体部脂肪组织的自动分割和定量测量研究  

Deep learning for automatic segmentation and quantification of body adipose tissue on MR images

作　　者：王可[1] 姜原[1] 黄嘉豪 王祥鹏 张晓东[1] 王霄英[1] WANG Ke;JIANG Yuan;HUANG Jia-hao(Department of Radiology,Peking University First Hospital 100034,China)

机构地区：[1]北京大学第一医院医学影像科,北京100034 [2]北京赛迈特锐医疗科技有限公司,100011

出　　处：《放射学实践》2021年第6期792-798,共7页Radiologic Practice

基　　金：2021年北大医学青年科技创新培育基金(BMU2021PYB021)。

摘　　要：目的:研究应用基于深度学习的3D-Unet网络模型在MRI图像中分割体部脂肪组织并生成定量结果的可行性。方法:回顾性搜集2020年4月1日-8月5日本院体部MR(包含GRE DIXON序列脂像)中符合入组标准的扫描病例用于模型训练。共搜集53例患者67个数据,包括胸部、腹部、盆腔图像各17、26、24例。由2名影像医生行图像标注,先以阈值分割方法将图像二值化,将脂肪组织分为皮下、肌骨、内脏3个区域,手工标注皮下脂肪、内脏脂肪得到标签。训练3D U-Net模型时将67个数据随机分为训练集(n=52)、调优集(n=6)和测试集(n=9)用于模型建立与评估。通过Dice系数、影像科医师评分来评价分割结果。根据分割结果生成脂肪体积、平均脂肪体积、脂肪比例、体部平均径线等结果,自动导入到结构化报告中。应用Wilcoxon配对检验、Pearson相关性分析、Bland-Altman分析、组内相关系数(ICC)将医师手工标注结果与模型输出结果比较。结果:内脏/皮下脂肪组织在模型训练集、调优集、测试集Dice系数分别为0.89/0.94;0.89/0.95和0.90/0.95。模型预测及手工标注内脏/皮下脂肪组织输出图像主观评分无统计学差异(P>0.05)。各部位模型预测结果生成脂肪体积、平均脂肪体积、脂肪比例、体部径线与医生手工标注结果之间的Pearson系数为0.968-1,ICC值为0.982-1,Bland-Altman分析显示良好的一致性。结论:MR图像基于深度学习行体部脂肪组织自动分割和定量测量可在技术上实现并有可能进一步研究此模型的临床应用价值。Objective:To develop a deep convolutional neural network to automatically segment MR images of the body for the measurement of adipose tissue.Methods:We manually segmented 67 MR images of the thorax,abdomen and pelvis in 53 subjects,labeling subcutaneous adipose tissue(AT),muscular-skeletal AT and visceral AT.All images underwent preprocessing with threshold segmentation.The data were randomly divided into training(n=52),validation(n=6)and test(n=9)datasets.The dice score was used to assess similarity between the manual segmentations and the 3D U-Net predicted segmentations.A subjective score given by radiologist was used to assess the quality of segmentations.The following data were obtained from the predicted labels,including total volume,average volume,and ratios of AT volume between subcutaneous to visceral area.All the quantitative results generated by AI model were automatically inputted to the structured report and compared to manual segmentations as the reference standard.Wilcoxon test,pearson test,Bland-Altman plot and ICC score were used to compare the difference of segmentation quality and data of the two methods.Results:The 3D U-Net model achieved accurate segmentation of AT for all classes.The dice scores for subcutaneous and visceral AT in different datasets were as follows:train(0.89,0.94),validate(0.89,0.95),and test(0.90,0.95).There was no difference of the subjective score given by radiologist between two segmentation methods.The quantitative parameters from AI model were not significantly different from that of the manual labeled results.Conclusions:The 3D U-Net model enables accurate and automatic segmentation of adipose tissue on body MR images,with promising implications for quantitative imaging studies.

关 键 词：磁共振成像 深度学习 脂肪组织 报告 

分 类 号：R445.2[医药卫生—影像医学与核医学] R-056[医药卫生—诊断学]