基于对抗图自编码的阿尔兹海默症脑网络分析

Adversarial Graph Autoencoder for Brain Network Analysis in Alzheimer’s Disease

阿尔兹海默症(Alzheimer’s Disease, AD)的不同阶段会发生结构或功能连接的改变。这些基于连接的特征可以大大提高疾病诊断的准确性,并能给出疾病的成因解释。如何有效融合结构和功能影像来挖掘不同模态之间的互补信息仍然是一个挑战。本文提出了一种对抗图自编码器模型,来提取脑连接特征用于AD分析。具体地说,将扩散张量成像(Diffusion Tensor Imaging, DTI)和功能磁共振成像(functional Magnetic Resonance Imaging, fMRI)相结合,构建每个受试者的图结构数据。图编码器(生成器)将图数据转换为潜在表征。同时,利用fMRI数据估计潜在分布,对图编码器进行正则化约束,以保证良好的潜在表征。为了保证潜在表征的稳定性,图解码器从潜在表征中恢复图数据。最后,将潜在表征送给分类器,使其具有疾病类别信息。实验结果表明,该模型比其他相关模型具有更高的预测精度。总体而言,该方法可以重建AD早期的结构–功能连接,分析异常的脑连接并用于AD的早期诊疗研究。

Alterations in the structural or functional connectivity take place at different stages of Alzheimer’s Disease (AD). These connectivity-based features can greatly improve the disease diagnosis accuracy and explain the causes of the disease. How to effectively fuse structural and functional images for exploring complementary information remains challenging. This paper proposes an adversarial graph autoencoder model to extract connectivity-based features for AD analysis. Specifically, Diffusion Tensor Imaging (DTI) and functional Magnetic Resonance Imaging (fMRI) are combined to construct graph data for each subject. The graph encoder (generator) transforms the graph data into a latent representation. Meanwhile, the fMRI data is utilized to estimate the latent distribution, which can regularize the graph encoder to ensure good latent representation. To ensure the latent representation is stable, the graph decoder regains the graph data from the latent representation. Finally, the latent representation is sent to the classifier to make it class-discriminative. Experimental results demonstrate that the proposed model can achieve higher prediction accuracy than other related models. Generally, this method can reconstruct the structural-functional connectivity and analyze abnormal brain connections for early AD study.