不同深度学习网络在心脏磁共振图像自动分析中的性能研究

Study of deep learning architectures on automatic analysis for cardiac parametric mapping

目的 深度学习具有端对端的预测能力,能够加速医学图像的后处理,但是目前尚缺乏比较不同神经网络在心脏磁共振定量图像自动化分析性能差异的研究。因此本研究旨在探究不同神经网络是否具有自动化心脏定量图像分析的能力并比较其性能。方法 本研究收集了155名健康志愿者的MOLLI、SASHA T_(1)定量图像和T_(2)-prep bSSFP T_(2)定量图像,分别对AlexNet、GoogLeNet、ResNet和DenseNet进行训练,使这些网络能够直接从T_(1)和T_(2)定量图像中预测左心室全局(左心室心肌和左心室血液)与局部(室间隔和AHA节段)T_(1)和T_(2)定量结果,并对其性能进行比较。结果 4种神经网络均能直接从心脏磁共振定量图像中预测全心和局部心肌的T_(1)和T_(2),具有自动化分析的能力。不同的网络对T_(1)和T_(2)的预测性能存在差异,同时对序列也敏感。在3个数据集中,DenseNet对于MOLLI和SASHA T_(1)定量图像的预测误差为17 ms±60 ms,对于T_(2)定量图像的预测误差约2 ms,具有最好的性能。GoogLeNet性能优于AlexNet和ResNet。结论 本研究实验证明神经网络能够自动化分析心脏磁共振定量图像,不同神经网络存在性能差异,本研究可为心脏磁共振图像自动化分析技术提供支持,为医生等提供一种快速的图像分析方法。

Objective Neural network with end-to-end prediction ability could accelerate medical imaging processing.Currently,there is no study to compare the difference in automatically analyzing cardiac parametric mapping between different neural networks.Therefore,this study aims to investigate the capability of various neural networks for automated cardiac parametric mapping analysis and to compare their performance.Methods This study collected MOLLI T_(1) images,SASHA T_(1) images,and T_2-prep bSSFP T_(2) images from 155 healthy volunteers and built four end-to-end neural networks with AlexNet,GoogLeNet,ResNet and DenseNet structures.We trained these neural networks to directly predict global [left ventricle(LV) myocardium and LV blood] and regional(ventricular septal and AHA segment) T_(1) and T_(2) values from corresponding maps.The strategies of training,validation,and testing for four networks were kept consistent.We compared their prediction performance.Results All four neural networks could successfully predict T_(1) and T_(2) for both global and region LV myocardium and LV blood with different accuracy,achieving automatic analysis.The performance of four networks also sensitivities to mapping sequence.DenseNet had the minimal residual error among all approaches.The prediction error of T_(1) for MOLLI and SASHA T_(1) maps was 17 ms±60 ms,and about 2 ms for T_(2) maps.The performance of GoogLeNet was superior to AlexNet and ResNet.Conclusions This study demonstrated that neural networks could automatically analyze cardiac parametric maps.The performance was dependent on the structure of neural networks.This study would promote the development of automatic techniques for cardiovascular magnetic resonance and provide an automatic tool for clinical practice.