The evaluation of DUNE: a U-Net-based neural network to denoise multi-echo fMRI data.

Area of Science:

• Neuroimaging
• Machine Learning
• Biomedical Engineering

Background:

• Task-based functional MRI (fMRI) data is often contaminated by scanner and physiological noise.
• This noise complicates the identification of task-related Blood-Oxygen-Level-Dependent (BOLD) signals.
• Existing methods like multi-echo ICA (MEICA) and high temporal resolution single-echo fMRI aim to improve BOLD signal detection.

Purpose of the Study:

• To introduce a novel U-shaped convolutional neural network, termed DUNE, for denoising multi-echo fMRI data.
• To evaluate DUNE's performance as an alternative to MEICA.
• To compare the activation maps generated by DUNE with those from MEICA and high temporal resolution single-echo fMRI.

Main Methods:

• Two multi-echo fMRI experiments were conducted.
• DUNE, a U-shaped convolutional neural network, was applied to denoise the multi-echo fMRI data.
• Denoised data from DUNE were compared against MEICA denoising and single-echo fMRI experiments.

Main Results:

• DUNE successfully reduced noise in multi-echo fMRI data.
• The BOLD effects of interest were preserved comparably to MEICA and single-echo fMRI.
• Activation maps generated using DUNE showed comparable quality to established methods.

Conclusions:

• DUNE is a viable and effective method for denoising multi-echo fMRI data.
• The U-shaped convolutional neural network approach shows potential for enhancing functional MRI analysis.
• DUNE offers a promising alternative for improving the detection of BOLD responses in noisy fMRI datasets.