Improved Image Reconstruction and Diffusion Parameter Estimation Using a Temporal Convolutional Network Model of Gradient Distortions

Motivation: Nonlinear gradient trajectory errors make imaging and parameter mapping challenging, especially in noncartesian imaging sequences. In many cases severe distortions dramatically impact image quality. Goal(s): To develop a nonlinear model that can accurately predict gradient distortions. Approach: We use a temporal convolutional network trained on measured gradient waveforms to predict gradient system outputs, and incorporate these predictions into image reconstruction. Results: Using the nonlinear TCN model results in improved image quality and diffusion parameter estimation over linear system models in a multishot imaging sequence. Impact: Nonlinear gradient errors do dramatically impact image quality but may be remediated with an accurate nonlinear model. Having a more accurate model of gradient distortions may allow for greater flexibility in the gradient waveforms used in MRI.