Current techniques for background field removal (BGFR), essential for quantitative susceptibility mapping, leave residual background fields and inaccuracies near air-tissue interfaces. We propose a new deep learning method aiming for robust brain BGFR: we trained a 3D U-net with realistic simulated and in-vivo data augmented with spatial deformations. The network trained on synthetic data predicts accurate local fields when tested on synthetic data, (median RMSE = 49.5%), but is less accurate when tested on in-vivo data. The network trained and tested on in-vivo data performs better, suggesting our synthetic set did not fully capture the complexity found in vivo.
This abstract and the presentation materials are available to members only; a login is required.