多通道稀疏图变换网络用于早期阿尔茨海默病识别

Multi-Channel Sparse Graph Transformer Network for Early Identification of Alzheimer′s Disease

目前阿尔茨海默病(AD)尚无有效的防治方法,早期临床干预可以延缓其进展,改善预后。然而现有的方法只考虑从群体关系中学习到的神经影像学特征而不考虑被试的个体特征。本研究设计了一种新颖的多模态多通道稀疏图变换网络(MSGTN)以期实现早期AD识别。首先,获取并处理每个被试者的影像信息[如弥散张量成像(DTI)和功能磁共振成像(fMRI)等]以及其相应的非影像信息(如年龄、性别等);其次,利用局部加权聚类系数(LWCC)将功能信息和结构信息进行融合,并将已融合的多模态影像特征与受试者的性别和年龄信息相结合来构建稀疏图;最后,将构建的稀疏图输入所设计的MSGTN网络模型用于早期AD识别。从公共数据库ADNI上获得共170个受试者,其中38个晚期轻度认知障碍(LMCI)患者,44个早期轻度认知障碍(EMCI)患者,44个显著记忆下降(SMC)患者和44个正常对照(NC)。结果表明,SMC与NC的准确度为87.02%,EMCI与NC的准确度为87.40%,LMCI与NC的准确度为91.49%,SMC与EMCI的准确度为88.93%、SMC与LMCI的准确度为86.74%、EMCI与LMCI的准确度为92.12%。所提出的诊断模型不仅能够分析出NC与3种不同早期AD疾病状态,而且在3种不同早期AD疾病状态中也取得了优越的分类性能。

Currently,there is no effective treatment for Alzheimer's disease(AD).Clinical intervention via early diagnosis can delay the AD progression and improve its prognosis.However,the existing methods only consider the neuroimaging features learned from group relationships,not the individual characteristics of the subjects.In this work,we designed a novel multi-modal multi-channel sparse graph transformer network(MSGTN).Our proposed network model included two parts,they are multi-modal data optimization and multi-modal feature learning.Firstly,we acquired the image information(e.g.,diffusion tensor imaging(DTI)and functional magnetic resonance imaging(fMRI))and non-image information(e.g.,age and sex)of each subject.Secondly,we utilized locally weighted clustering coefficients(LWCC)to fuse functional and structural information.After that,the fused multi-modal image features were combined with the gender and age information of the subjects to construct a sparse graph.Finally,we input the sparse graph into the MSGTN network for early AD identification.We obtained a total of 170 subjects from the public database ADNI(Alzheimer's disease neuroimaging initiative),including 38 LMCI,44 EMCI,44 SMC,and 44 normal controls(NC).Our method achieved classification accuracy of 87.02%,87.40%,91.49%,88.93%,86.74%and 92.12%,respectively.The experimental results have proved that our proposed model not only can analyze NC versus three different early AD disease states,but also achieved superior classification performance in three different early AD disease states.