GDCNet: deep learning model for self-consistent geometric distortion correction of EPI images without a field map

GDCNet is a “self-consistent” deep learning (DL) model for distortion correction of EPI fMRI images and field map estimation. It only requires the EPI images for correction, saving acquisition time and avoiding motion-related correction errors. The two supervised U-Nets for forward modelling and distortion correction have been tested in silico and in vivo on a publicly-available and a prospectively-acquired dataset. The in silico models demonstrated generalization capabilities and achieved a mean RMSE of 2.56 x10-2 as self-conistency metric. Inference in vivo showed modest correction in the prefrontal cortex and similar estimated field map compared to the acquired ground truth.