A robust deep neural network for denoising task-based fMRI data: An application to working memory and episodic memory.

Abstract

In this study, a deep neural network (DNN) is proposed to reduce the noise in task-based fMRI data without explicitly modeling noise. The DNN artificial neural network consists of one temporal convolutional layer, one long short-term memory (LSTM) layer, one time-distributed fully-connected layer, and one unconventional selection layer in sequential order. The LSTM layer takes not only the current time point but also what was perceived in a previous time point as its input to characterize the temporal autocorrelation of fMRI data. The fully-connected layer weights the output of the LSTM layer, and the output denoised fMRI time series is selected by the selection layer. Assuming that task-related neural response is limited to gray matter, the model parameters in the DNN network are optimized by maximizing the correlation difference between gray matter voxels and white matter or ventricular cerebrospinal fluid voxels. Instead of targeting a particular noise source, the proposed neural network takes advantage of the task design matrix to better extract task-related signal in fMRI data. The DNN network, along with other traditional denoising techniques, has been applied on simulated data, working memory task fMRI data acquired from a cohort of healthy subjects and episodic memory task fMRI data acquired from a small set of healthy elderly subjects. Qualitative and quantitative measurements were used to evaluate the performance of different denoising techniques. In the simulation, DNN improves fMRI activation detection and also adapts to varying hemodynamic response functions across different brain regions. DNN efficiently reduces physiological noise and generates more homogeneous task-response correlation maps in real data.