Correction of Resting-State fMRI Data Contaminated by B0 Inhomogeneity Artifacts Using a Single-Subject Fitting Neural Network

Motivation: A primary limitation of Gradient Echo-EPI(GRE-EPI) is its sensitivity to B0 inhomogeneity, resulting in signal dropout at air/tissue interfaces. Current artifact correction methods, however, continue to face substantial limitation. Goal(s): We aim to develop a frequency domain based single subject fitting neural network to effectively reduce B0 field inhomogeneity artifacts. Approach: Spin Echo EPI is resilient to B0 inhomogeneity but less T2*-sensitive. GRE EPI is T2*-sensitive but prone to B0 inhomogeneity. Leveraging these properties, we proposed a single-subject fitting neural network. Results: The proposed method mitigated the B0 filed inhomogeneity artifacts and restored brain signals. Impact: B0 field inhomogeneity induced artifacts were mitigated by proposed frequency-domain deconvolutional neural network. The proposed method is expected to broaden the detectable range of brain regions, particularly in areas heavily affected by B0 inhomogeneity, enabling more comprehensive whole-brain fMRI research.