Whole Brain fMRI Pattern Analysis Based on Tensor Neural Network

Functional magnetic resonance imaging (fMRI) has increasingly come to dominate brain mapping research, as it provides a dynamic view of brain matter. Feature selection or extraction methods play an important role in the successful application of machine learning techniques to classifying fMRI data by appropriately reducing the dimensionality of the data. While whole-brain fMRI data contains large numbers of voxels, the curse of dimensionality problem may limit the feature selection/extraction and classification performance of traditional methods. In this paper, we propose a novel framework based on a tensor neural network (TensorNet) to extract the essential and discriminative features from the whole-brain fMRI data. The tensor train model was employed to construct a simple and shallow neural network and compress a large number of network weight parameters. The proposed framework can avoid the curse of dimensionality problem, and allow us to extract effective patterns from the whole-brain fMRI data. Furthermore, it reveals a new perspective for analyzing complex fMRI data with a large numbers of voxels, through compressing the number of parameters in a neural network. Experimental results confirmed that our proposed classification framework based on TensorNet outperforms traditional methods based on an SVM classifier for multi-class fMRI data.