融合生成对抗网络的青少年脑MRI配准体测研究

A body examination study of brain MRI registration in adolescents using fusion generative adversarial network

目的为了解决UNet框架上采样过程中信息丢失的问题,本文采用青少年脑部MRI研究网络学习能力弱和脑部边缘区域配准精度不高的问题。材料与方法本文采用公开可用的脑部MRI数据集:HBN和LPBA40,提出了一种结合多尺度注意力机制的生成对抗网络(multiscale attention mechanisms generative adversarial networks,MAMGAN)配准模型,实现了单模态脑图像配准。该方法由配准网络和鉴别网络组成。通过在鉴别网络中添加多尺度注意力机制(multiscale attention mechanisms,MAM)模块获取不同尺度下的上下文信息,在对抗训练过程中提取到更有效的大脑结构特征。其次,在配准网络中引入了图像相似性的局部互相关损失函数,约束移动图像与固定图像之间的相似性,在两个网络的对抗训练过程中进一步提高图像配准的性能。本文使用Dice系数(Dice coefficient,Dice)、结构相似度(structural similarity,SSIM)和皮尔森相关系数(Pearson’s correlation coefficient,PCC)衡量配准图像与固定图像的配准精度。结果MAM GAN方法在Dice指标上相对于传统的方法,脑脊液(cerebrospinal fluid,CSF)、脑灰质(gray matter,GM)和脑白质(white matter,WM)精度分别提高了0.013、0.023和0.028,PCC指标提高了0.004,SSIM指标提高了0.011。由此可见,该方法配准效果好。结论MAM GAN方法能够更好地学习到脑部结构特征,提升了配准的性能,为青少年多动症临床诊断和体质检测提供技术基础。

Objective:To solve the problem of information loss in the sampling process of UNet framework,we used brain MRI of adolescents to study the problems of weak network learning ability and low accuracy of registration of brain marginal regions.Materials and Methods:In this study,publicly available brain MRI data sets were used:HBN and LPBA40 propose a multiscale attention mechanisms generative adversarial networks(MAM_GAN).Single-mode brain image registration was realized.The method consists of registration network and authentication network.By adding multiscale attention mechanisms(MAM)modules to the identification network to acquire contextual information at different scales,more effective brain structural features were extracted during adversarial training.Secondly,the local cross-correlation loss function of image similarity was introduced into the registration network to constrain the similarity between the moving image and the fixed image,which further improves the image registration performance in the antagonistic training process of the two networks.Dice coefficient(Dice),structural similarity(SSIM)and Pearson’s correlation coefficient(PCC)were used to measure the registration accuracy of registration image and fixed image.Results:Compared with the traditional methods in Dice score,the accuracy of MAM_GAN method in cerebrospinal fluid(CSF),gray matter(gray matter,GM)and white matter(white matter,WM)increased by 0.013,0.023 and 0.028 respectively,PCC score increased by 0.004 and SSIM score increased by 0.011.Hence,the experimental results showed that the method had good registration effect.Conclusions:The MAM_GAN method can better learn the structural features of the brain,improve the registration performance,and provide a technical basis for the clinical diagnosis and physical detection of attention-deficit hyper-activity disorder(ADHD)in adolescents.