BrainGNN: Interpretable Brain Graph Neural Network for fMRI Analysis

Abstract

The brain is an exceptionally complex system and understanding it's functional 2 organization is the goal of modern neuroscience. Using fMRI, large strides in 3 understanding this organization have been made by modeling the brain as a 4 graph-a mathematical construct describing the connections or interactions (i.e. 5 edges) between different discrete objects (i.e. nodes). To create these graphs, 6 nodes are defined as brain regions of interest (ROIs) and edges are defined as the 7 functional connectivity between those ROIs, computed as the pairwise correla-8 tions of functional magnetic resonance imaging (fMRI) time series, as illustrated 9 in Fig. 1. Traditional graph-based analyses for fMRI have focused on using graph 10 theoretical metrics to summarize the functional connectivity for each node into 11 a single number [45,23]. However, these methods do not consider higher-order 12 interactions between ROIs, as these interactions cannot be preserved in a sin-13 gle number. Additionally, due to the high dimensionality of fMRI data, usually 14 ROIs are clustered into highly connected communities to reduce dimensionality. 15Then, features are extracted from these smaller communities for further analysis 16 [31,12]. For these two-stage methods, if the results from the first stage are not 17 reliable, significant errors can be induced in the second stage. 18The past few years have seen the growing prevalence of the use of graph 19 neural networks (GNN) for end-to-end graph learning applications. GNNs are 20 the state-of-the-art deep learning methods for most graph-structured data anal-21 ysis problems. They combine node features, edge features, and graph structure 22 by using a neural network to embed node information and pass information 23 through edges in the graph. As such, they can be viewed as a generalization of 24 the traditional convolutional neural networks (CNN) for images. Due to their 25 high performance and interpretability, GNNs have been a widely applied graph 26 analysis method. [26,25,49,28,50]. Most existing GNNs are built on graphs that 27 do not have correspondence between the nodes of different instances, such as 28 social networks and protein networks, limiting interpretability. These methods 29-including the current GNN methods for fMRI analysis -use the same ker-30 nel over different nodes, which implicitly assumes brain graphs are translation 31 invariant. However, nodes in the same brain graph have distinct locations and 32 unique identities. Thus, applying the same kernel over all nodes is problematic. 33In addition, few GNN studies have explored both individual-level and group-level 34 explanations, which are critical in neuroimaging research. 35In this work, we propose a graph neural network-based framework for map-36 ping regional and cross-regional functional activation patterns for classification 37 tasks, such as classifying neurodisorder patients versus healthy control subjects 38 and performing cognitive task decoding. Our framework jointly learns ROI clus-39 tering and the downstr...

Graph-Based Deep Learning for Medical Diagnosis and Analysis: Past, Present and Future

“…In other words, the GCN is trained on the whole graph and tested on sub-graphs, such that they could determine the importance of sub-graphs and nodes. In both works from Li et al [ 17 , 61 ], the authors also improved their individual graph level analysis by proposing a BrainGNN and a pooling regularized GNN model to investigate the brain region related to a neurological disorder from t- f MRI data for ASD or HC classification.…”

“…Thus, applying the same kernel over all nodes is problematic. Li et al [ 61 ] adopted weighted graphs from f MRI and ROI-aware graph convolutional layers to infer which ROIs are important for prediction of cognitive tasks. The model maps regional and cross-regional functional activation patterns for classification of cognitive task decoding in the HCP 900 dataset [ 76 ].…”

“…Post-hoc explanation techniques have been used to visualize attention weights or to recognize relevant subgraphs for classification. Such explanation techniques have been useful, for example, to identify brain regions that are most involved when a seizure occurs [ 59 , 86 ], and key brain regions associated with biomarkers for PD [ 47 ] or ASD [ 16 , 17 , 61 ] prediction. The activation map and gradient sensitivity of graph models are also used to interpret the salient input features at both the group and individual levels for the analysis of BD [ 44 ] and ASD [ 61 ].…”

A Deep Graph Neural Network Architecture for Modelling Spatio-temporal Dynamics in resting-state functional MRI Data

“…Our architecture can very easily include other types of data for future work (e.g., multimodal structural and temporal data), and be extended to include possible confounds that could drive the prediction task in other brain disorders. Another exciting recent trend that can be included in our architecture is to allow the network to learn the underlying connectivity from scratch [48, 49] instead of computing associations or other handcrafted features like the ones used in this and other works [50, 51].…”

Generalizable Machine Learning in Neuroscience Using Graph Neural Networks

“…With the success of convolutional neural networks, researchers successfully applied deep learning to numerous domains in neuroscience Glaser et al (2019) including MRI imaging Lundervold and Lundervold (2019) and connectomes Brown and Hamarneh (2016) where algorithms can predict disorders such as autism Brown et al (2018) . Further leveraging the explicit graph structure of neural systems, several studies have successfully applied GNNs on various tasks such as annotating cognitive state Zhang and Bellec, 2019 , and several frameworks based on graph neural networks have been proposed for analyzing fMRI data [ Li and Duncan (2020) ; Kim and Ye (2020) ].…”