Introducing ME-DUNE, a denoising U-network applied in a task based multi-echo fMRI study.

Motivation: fMRI data is noisy and the standard denoising technique used in multi-echo (ME) fMRI is ICA-based (ME-ICA). Consequently, its performance depends on how well ICA is able to separate noise from non-noise components. Goal(s): To evaluate the potential of using an U-convolutional network to denoise ME-fMRI data. Approach: The efficacy of our denoising U-network (DUNE) was compared to ME-ICA in task-based ME-fMRI by looking at the residual noise and the activation maps. Results: DUNE was found to be effective in reducing the noise while preserving the BOLD response of interest while ME-ICA failed to denoise the data. Impact: As a first step in the development of a new denoising technique that is not ICA-dependent, this work showed the potential of using an U-convolutional network to denoise multi-echo fMRI data as an alternative to ICA-based methods.