Application of Convolutional Recurrent Neural Network for Individual Recognition Based on Resting State fMRI Data

A transformer model for learning spatiotemporal contextual representation in fMRI data

Representation learning is a core component in data-driven modeling of various complex phenomena. Learning a contextually informative representation can especially benefit the analysis of fMRI data because of the complexities and dynamic dependencies present in such datasets. In this work, we propose a framework based on transformer models to learn an embedding of the fMRI data by taking the spatiotemporal contextual information in the data into account. This approach takes the multivariate BOLD time series of the regions of the brain as well as their functional connectivity network simultaneously as the input to create a set of meaningful features that can in turn be used in various downstream tasks such as classification, feature extraction, and statistical analysis. The proposed spatiotemporal framework uses the attention mechanism as well as the graph convolution neural network to jointly inject the contextual information regarding the dynamics in time series data and their connectivity into the representation. We demonstrate the benefits of this framework by applying it to two resting-state fMRI datasets, and provide further discussion on various aspects and advantages of it over a number of other commonly adopted architectures.

Representation learning of resting state fMRI with variational autoencoder

Resting state functional magnetic resonance imaging (rsfMRI) data exhibits complex but structured patterns. However, the underlying origins are unclear and entangled in rsfMRI data. Here we establish a variational auto-encoder, as a generative model trainable with unsupervised learning, to disentangle the unknown sources of rsfMRI activity. After being trained with large data from the Human Connectome Project, the model has learned to represent and generate patterns of cortical activity and connectivity using latent variables. The latent representation and its trajectory represent the spatiotemporal characteristics of rsfMRI activity. The latent variables reflect the principal gradients of the latent trajectory and drive activity changes in cortical networks. Representational geometry captured as covariance or correlation between latent variables, rather than cortical connectivity, can be used as a more reliable feature to accurately identify subjects from a large group, even if only a short period of data is available in each subject. Our results demonstrate that VAE is a valuable addition to existing tools, particularly suited for unsupervised representation learning of resting state fMRI activity.

Key Intrinsic Connectivity Networks for Individual Identification With Siamese Long Short-Term Memory

In functional magnetic resonance imaging (fMRI) analysis, many studies have been conducted on inter-subject variability as well as intra-subject reproducibility. These studies indicate that fMRI could have unique characteristics for individuals. In this study, we hypothesized that the dynamic information during 1 min of fMRI was unique and repetitive enough for each subject, so we applied long short-term memory (LSTM) using initial time points of dynamic resting-state fMRI for individual identification. Siamese network is used to obtain robust individual identification performance without additional learning on a new dataset. In particular, by adding a new structure called region of interest–wise average pooling (RAP), individual identification performance could be improved, and key intrinsic connectivity networks (ICNs) for individual identification were also identified. The average performance of individual identification was 97.88% using the test dataset in eightfold cross-validation analysis. Through the visualization of features learned by Siamese LSTM with RAP, ICNs spanning the parietal region were observed as the key ICNs in identifying individuals. These results suggest the key ICNs in fMRI could represent individual uniqueness.

Decoding Multiple Sound-Categories in the Auditory Cortex by Neural Networks: An fNIRS Study

This study aims to decode the hemodynamic responses (HRs) evoked by multiple sound-categories using functional near-infrared spectroscopy (fNIRS). The six different sounds were given as stimuli (English, non-English, annoying, nature, music, and gunshot). The oxy-hemoglobin (HbO) concentration changes are measured in both hemispheres of the auditory cortex while 18 healthy subjects listen to 10-s blocks of six sound-categories. Long short-term memory (LSTM) networks were used as a classifier. The classification accuracy was 20.38 ± 4.63% with six class classification. Though LSTM networks' performance was a little higher than chance levels, it is noteworthy that we could classify the data subject-wise without feature selections.

Functional connectome fingerprinting using shallow feedforward neural networks

Although individual subjects can be identified with high accuracy using correlation matrices computed from resting-state functional MRI (rsfMRI) data, the performance significantly degrades as the scan duration is decreased. Recurrent neural networks can achieve high accuracy with short-duration (72 s) data segments but are designed to use temporal features not present in the correlation matrices. Here we show that shallow feedforward neural networks that rely solely on the information in rsfMRI correlation matrices can achieve state-of-the-art identification accuracies (≥99.5%) with data segments as short as 20 s and across a range of input data size combinations when the total number of data points (number of regions × number of time points) is on the order of 10,000.

Graph convolutional network for fMRI analysis based on connectivity neighborhood

There have been successful applications of deep learning to functional magnetic resonance imaging (fMRI), where fMRI data were mostly considered to be structured grids, and spatial features from Euclidean neighbors were usually extracted by the convolutional neural networks (CNNs) in the computer vision field. Recently, CNN has been extended to graph data and demonstrated superior performance. Here, we define graphs based on functional connectivity and present a connectivity-based graph convolutional network (cGCN) architecture for fMRI analysis. Such an approach allows us to extract spatial features from connectomic neighborhoods rather than from Euclidean ones, consistent with the functional organization of the brain. To evaluate the performance of cGCN, we applied it to two scenarios with resting-state fMRI data. One is individual identification of healthy participants and the other is classification of autistic patients from normal controls. Our results indicate that cGCN can effectively capture functional connectivity features in fMRI analysis for relevant applications.

A New Spiking Convolutional Recurrent Neural Network (SCRNN) With Applications to Event-Based Hand Gesture Recognition

The combination of neuromorphic visual sensors and spiking neural network offers a high efficient bio-inspired solution to real-world applications. However, processing event- based sequences remains challenging because of the nature of their asynchronism and sparsity behavior. In this paper, a novel spiking convolutional recurrent neural network (SCRNN) architecture that takes advantage of both convolution operation and recurrent connectivity to maintain the spatial and temporal relations from event-based sequence data are presented. The use of recurrent architecture enables the network to have a sampling window with an arbitrary length, allowing the network to exploit temporal correlations between event collections. Rather than standard ANN to SNN conversion techniques, the network utilizes a supervised Spike Layer Error Reassignment (SLAYER) training mechanism that allows the network to adapt to neuromorphic (event-based) data directly. The network structure is validated on the DVS gesture dataset and achieves a 10 class gesture recognition accuracy of 96.59% and an 11 class gesture recognition accuracy of 90.28%.