Integrating Multimodal MRIs for Adult ADHD Identification with Heterogeneous Graph Attention Convolutional Network

Abstract

Adult attention-deficit/hyperactivity disorder (ADHD) is a mental health disorder whose symptoms would change over time. Compared with subjective clinical diagnosis, objective neuroimaging biomarkers help us better understand the mechanism of brain between patients with brain disorders and age-matched healthy controls. In particular, different magnetic resonance imaging (MRI) techniques can depict the brain with complementary structural or functional information. Thus, effectively integrating multi-modal MRIs for ADHD identification has attracted increasing interest. Graph convolutional networks (GCNs) have been applied to model brain structural/functional connectivity patterns to discriminate mental disorder from healthy controls. However, existing studies usually focus on a specific type of MRI, and therefore cannot well handle heterogeneous multimodal MRIs. In this paper, we propose a heterogeneous graph attention convolutional network (HGACN) for ADHD identification, by integrating resting-state functional MRI (fMRI) and diffusion MRI (dMRI) for a comprehensive description of the brain. In the proposed HGACN, we first extract features from multimodal MRI, including functional connectivity for fMRI and fractional anisotropy for dMRI. We then integrate these features into a heterogeneous brain network via different types of metapaths, with each type of metapatch corresponding to a specific modality or functional/structural relationship between regions of interest. We leverage both intra-metapth and inter-metapath attention to learn useful graph embeddings/representations of multimodal MRIs and finally predict subject’s category label using these embeddings. Experimental results on 110 adult ADHD patients and 77 age-matched HCs suggest that our HGACN outperforms several state-of-the-art methods in ADHD identification based on resting-state functional MRI and diffusion MRI.