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

Task based fMRI data suffers from scanner and physiologic noise. Consequently, finding the task based BOLD responses out of the noise is challenging. To improve the power to detect the BOLD responses, multi-echo (ME) fMRI combined with ICA based denoising (MEICA) and single-echo (SE) fMRI at high temporal resolution (<1 s) have been introduced. Both techniques have been found to give better activation maps than a traditional fMRI experiment at low temporal resolution (1.5-3 s). In this study, we introduced a new U-shaped convolutional neural network DUNE to denoise ME-fMRI data as an alternative to MEICA in 2 ME-fMRI experiments. The resulting activation maps found after denoising with DUNE were compared with those found after denoising with MEICA and similar SE-fMRI experiments at high temporal resolution. Our results revealed that DUNE was successful in reducing the noise while preserving the BOLD effects of interest comparable to MEICA and SE-fMRI. This latter result showed the potential of using a U-shaped convolutional neural network DUNE to denoise ME-fMRI data.