基于交叉双流特征融合配准网络对阿尔茨海默病中大脑皮质及皮下核团的图像分析  

Image analysis of cortical and subcortical nuclei in Alzheimer disease based on intersected dual stream feature fusion registration network

作　　者：李振宇 李恩慧 张童禹 张唯唯 LI Zhenyu;LI Enhui;ZHANG Tongyu;ZHANG Weiwei(Institute of Basic Medical Sciences,Chinese Academy of Medical Sciences&Peking Union Medical College,Beijing 100005)

机构地区：[1]中国医学科学院基础医学研究所,北京协和医学院基础学院,北京100005

出　　处：《北京生物医学工程》2025年第1期16-25,共10页Beijing Biomedical Engineering

基　　金：中国医学科学院医学与健康科技创新工程项目(CIFMS 2021-I2M-1-025)资助。

摘　　要：目的提出一种基于交叉双流的多尺度注意力特征融合网络(命名为MAFF-Net),用于脑图像微分同胚配准,以实现阿尔茨海默病(Alzheimer disease,AD)相关的脑结构标签的快速提取和分析。方法首先利用交叉双流网络推断图像对之间的相互映射关系,并通过引入注意力机制融合多尺度特征信息;然后利用微分同胚配准增强形变场的连续性和全局平滑性提高配准质量;最后,在自采集、OASIS-AD与OASIS-Health数据集上进行脑图像配准实验,采用Dice相似性系数(Dice similarity coefficients,DSC)、召回率(recall)、平均表面距离(average surface distance,ASD)及雅克比行列式(Jacobian determinant)验证MAFF-Net模型的性能,并进一步分析OASIS数据集的脑结构标签提取结果。结果脑图像配准实验结果显示,MAFF-Net算法在三个测试集上DSC分别为0.832、0.853和0.865,负雅可比行列式体素比例分别为0.027%、0.192%和0.089%,召回率分别为0.924、0.909和0.920,ASD分别为0.447 mm、0.387 mm和0.345 mm,除召回率外其余指标均优于对比算法。OASIS数据集的脑结构标签分析结果表明,大脑皮质、海马体和杏仁核的体积和表面积与年龄和健康状态存在密切联系。结论本文提出的MAFF-Net模型可以获得脑MR图像精确的配准性能和标签提取结果,通过AD相关的脑结构形态学特征分析,为AD早期诊断提供辅助参考价值。Objective An attention-based multiscale feature fusion network with intersected dual stream was proposed,namely MAFF-Net,for diffeomorphic brain image registration,in order to achieve rapid extraction and analysis of Alzheimer disease-related brain structure labels.Methods The intersected dual stream network was used to infer the mutual mapping relationship between image pairs,then the multiscale feature information was fused by introducing the attention mechanism,finally diffeomorphic registration was introduced to enhance the continuity and global smoothness of the deformation field and improve the registration quality.Brain image registration experiments were conducted on self-collected,OASIS-AD,and OASIS-Health datasets.The performance of the MAFF-Net model was validated using metrics by Dice similarity coefficient(DSC),recall,average surface distance(ASD),and the Jacobian determinant.Further analysis was performed on the brain structure label extraction results from the OASIS dataset.Results The experimental results of brain image registration showed that the MAFF-Net algorithm had DSC values of 0.832,0.853,and 0.865 on the three test sets,negative Jacobian determinant voxel ratios of 0.027%,0.192%,and 0.089%,recall values of 0.924,0.909,and 0.920,ASD values of 0.447 mm,0.387 mm,and 0.345 mm,with all but recall being superior to the comparison algorithm.The results of brain structural label analysis on the OASIS dataset showed that the volume and surface area of the cerebral cortex,hippocampus,and amygdala were closely related to age and health status.Conclusions The MAFF-Net model proposed in this paper can obtain accurate registration performance and label extraction results of brain MR Images,and provide auxiliary reference value for the early diagnosis of AD through the analysis of morphological characteristics of AD related brain structures.

关 键 词：MR脑图像 微分同胚配准 注意力机制 阿尔茨海默病 脑解剖结构 

分 类 号：R318.04[医药卫生—生物医学工程]