Denoising Task-Based Functional Magnetic Resonance Imaging Data Using Hybrid Deep Neural Network Model

Abstract

We present a hybrid deep neural network model for Denoising task-based fMRI data. This model is based on the Deep Neural Network (DNN) model [1] and improves its performance by utilizing different layers for Grey Matter and non-Grey Matter related data. The preprocessed data passes through GM and non-GM masks. Then it is fed to the network and passes through temporal filters along with an LSTM layer. The result will be classified using time-distributed fully-connected layers –since the input data is a time-series of voxel values-which have different structures for GM and non-GM input data. This hybrid mechanism helps the model to reduce the correlation between non-GM processed data and the design matrix significantly. We also increase the correlation of GM output, which is equivalent to better cost function and less noise. This new self-supervised model is tested using simulated data and the HCP database [2]. We succeed in denoising task better than the DNN model by an average of at least 4%.