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

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Signal Processing

Background:

• Task-based functional magnetic resonance imaging (fMRI) is susceptible to various noise sources that can obscure neural signals.
• Traditional denoising methods often require explicit modeling of specific noise types, which can be limiting.

Purpose of the Study:

• To propose and evaluate a deep neural network (DNN) for denoising task-based fMRI data without explicit noise modeling.
• To improve the detection of task-related neural responses and the quality of fMRI data.

Main Methods:

• A DNN architecture comprising temporal convolutional, LSTM, fully-connected, and selection layers was developed.
• Model parameters were optimized by maximizing the signal-to-noise ratio difference between gray matter and non-gray matter voxels.
• The DNN was applied to simulated fMRI data, working memory task data, and episodic memory task data.

Main Results:

• The DNN demonstrated improved fMRI activation detection in simulations, adapting to varied hemodynamic response functions.
• It effectively reduced physiological noise in real fMRI data.
• The DNN generated more homogeneous task-response correlation maps compared to traditional methods.

Conclusions:

• The proposed DNN offers a powerful and adaptable approach for denoising task-based fMRI data.
• This method enhances the extraction of task-related signals and improves the reliability of neuroimaging findings.