Introducing a task design independent customized cost function for DNN to denoise task-based fMRI data.

Motivation: We expected that using a fMRI denoising neural network (DNN) that requires the denoised signal to correlate with a task design matrix in combination with a GLM analysis, can lead to biased results. Goal(s): To find a DNN cost function that is independent of the task design. Approach: We suggested a cost function that is based on the preserved frequency content and the correlation with motion regressors, with non-brain signals and with the non-denoised signal. Results: We found that the DNN with the proposed cost function performed best in reducing the noise while preserving the BOLD signal. Impact: Our intension with this abstract is to make researchers to think more carefully about the conditions included in their customized cost function in DNN-like denoising models and to check critically the denoised output.