Deep neural networks on graph signals for brain imaging analysis

Abstract

Brain imaging data such as EEG or MEG are high-dimensional spatiotemporal data often degraded by complex, non-Gaussian noise. For reliable analysis of brain imaging data, it is important to extract discriminative, low-dimensional intrinsic representation of the recorded data. This work proposes a new method to learn the lowdimensional representations from the noise-degraded measurements. In particular, our work proposes a new deep neural network design that integrates graph information such as brain connectivity with fully-connected layers. Our work leverages efficient graph filter design using Chebyshev polynomial and recent work on convolutional nets on graph-structured data. Our approach exploits graph structure as the prior side information, localized graph filter for feature extraction and neural networks for high capacity learning. Experiments on real MEG datasets show that our approach can extract more discriminative representations, leading to improved accuracy in a supervised classification task.

Disease prediction using graph convolutional networks: Application to Autism Spectrum Disorder and Alzheimer’s disease

“…As an alternative, spectral methods have been used to learn a similarity metric between functional connectivity networks , which was applied for Autism Spectrum Disorder (ASD) and sex classification as well as manifold learning (Ktena et al, 2018). Spectral methods have also been explored for the prediction of visual tasks from MEG signals on a small number of subjects (Guo et al, 2017), while a bootstrapping strategy was used by Anirudh and Thiagarajan (2017) for ASD prediction. Finally, Lombaert et al (2015) combine spectral theory with random forests to process brain surfaces using spectral representations of meshes.…”

Data augmentation using generative adversarial neural networks on brain structural connectivity in multiple sclerosis

“…The performance metrics used for the evaluation are 1 , Precision and Accuracy which are defined in Eq. (11,12,13) respectively.…”

“…Recently, with the rise of DL, interesting alternatives have appeared and new generative DL-based models were proposed to obtain synthetic data with characteristics spanning the original data manifold [10]. Therefore, in this study we refer to generative models as a subclass of DL frameworks able to generate complex data data structure, including the recent modeling approach used to characterize brain networks by means of graph theory [11,12,13]. Given the great capability of graphs to represent complex relations among different areas of the brain, such relational data structure started to be widely employed in many contexts, including social behavioral studies.…”

Eeg-Based Video Identification Using Graph Signal Modeling and Graph Convolutional Neural Network

“…It was shown that several deep neural networks such as deep belief networks, stacked denoising autoencoders, and convolutional neural networks can extract effective application-driven features for EEG signals in spatial and spectral domains [10] [11][12] [13]. Deep learning models on graph signals, especially graph convolutional neural networks (GCNN) [5] [6] have been considered as competitive approaches for analyzing MEG [14] and fMRI [15] signals. However, graph signal-based deep learning for EEG has been rarely found in literature.…”

“…However, graph signal-based deep learning for EEG has been rarely found in literature. A challenge arises from the fact that EEG signals usually have smaller numbers of brain region representatives, i.e., electrodes (e.g., 32 in [16]), than MEG (e.g., 306 in [14]) or fMRI (e.g., 110 in [15]), so it is difficult to construct rich graph structures for EEG signals.…”