基于深度学习联合分割模型对扩张型心肌病心肌纤维化的定量分析

Quantitative analysis of myocardial fibrosis in dilated cardiomyopathy with deep learning joint segmentation model

目的探讨基于深度学习的心肌-纤维化区域联合分割模型对扩张型心肌病(DCM)患者心肌纤维化定量分析的效果。方法回顾性分析徐州市中心医院2015年1月至2022年4月确诊为DCM,并接受心脏MR-钆延迟强化检查显示左心室心肌纤维化的200例患者资料,分为训练集120例、验证集30例、测试集50例。由影像科医师勾勒左心室心肌轮廓和选取正常心肌区域,应用标准差法(SD)计算阈值提取纤维化心肌,作为左心室分割和纤维化量化的参考标准。通过凸形先验的U-Net网络分割左心室心肌,然后应用VGG图像分类网络识别正常心肌图像块,计算SD阈值提取纤维化心肌。采用精确度、召回率、交并比和Dice系数评价心肌分割效果。采用组内相关系数(ICC)评价深度学习联合分割模型与手动提取测得的左心室心肌纤维化比率的一致性。根据纤维化比率中位数,将测试集样本分为轻度组和重度组,通过Mann-Whitney U检验比较纤维化量化效果差异。结果在测试集中,心肌分割精确度为0.827(0.799,0.854),召回率为0.849(0.822,0.876),交并比为0.788(0.760,0.816),Dice系数为0.832(0.807,0.857)。联合分割模型与手动提取的纤维化比率的一致性高(ICC=0.991,P<0.001)。轻度和重度纤维化比率的误差率差异无统计学意义(P>0.05)。结论该联合分割模型实现了左心室心肌纤维化比率的自动计算,与医师手动提取结果一致性高,能够较为精准地实现DCM患者的心肌纤维化自动定量分析。

Objective To explore the effect of joint segmentation model of myocardial-fibrotic region based on deep learning in quantitative analysis of myocardial fibrosis in patients with dilated cardiomyopathy(DCM).Methods The data of 200 patients with confirmed DCM and myocardial fibrosis in the left ventricle detected by cardiac MR-late gadolinium enhancement(CMR-LGE)in Xuzhou Central Hospital from January 2015 to April 2022 were retrospectively analyzed.Using a complete randomized design,the patients were divided into training set(n=120),validation set(n=30)and test set(n=50).The left ventricle myocardium was outlined and the normal myocardial region was selected by radiologists.Fibrotic myocardium was extracted through calculating the threshold with standard deviation(SD)as a reference standard for left ventricle segmentation and fibrosis quantification.The left ventricular myocardium was segmented by convex prior U-Net network.Then the normal myocardial image block was recognized by VGG image classification network,and the fibrosis myocardium was extracted by SD threshold.The myocardial segmentation effect was evaluated using precision,recall,intersection over union(IOU)and Dice coefficient.The consistency of myocardial fibrosis ratio in left ventricle obtained by joint segmentation model and manual extraction was evaluated with intra-class correlation coefficient(ICC).According to the median of fibrosis rate,the samples were divided into mild and severe fibrosis,and the quantitative effect of fibrosis was compared by Mann-Whitney U test.Results In the test set,the precision of myocardial segmentation was 0.827(0.799,0.854),the recall was 0.849(0.822,0.876),the IOU was 0.788(0.760,0.816),and the Dice coefficient was 0.832(0.807,0.857).The consistency of fibrosis ratio between joint segmentation model and manual extraction was high(ICC=0.991,P<0.001).No statistically significant difference was found in the ratio error between mild and severe fibrosis(P>0.05).Conclusions The joint segmentation model realizes the automatic calculation of myocardial fibrosis ratio in left ventricle,which is highly consistent with the results of manual extraction.Therefore,it can accurately realize the automatic quantitative analysis of myocardial fibrosis in patients with dilated cardiomyopathy.