Deep Gated Convolutional Neural Network for QSM Background Field Removal

Abstract

Magnetic-resonance based quantitative susceptibility mapping (QSM) requires the removal of background magnetic perturbation fields for precise quantification of tissue magnetism within a volume of interest (VOI). Conventional state-of-the-art QSM background removal methods suffer from several limitations in accuracy and performance. To overcome these limitations, a 3D gated convolutional neural network was trained to infer whole brain local tissue magnetic perturbation fields, using total field perturbation measurements and brain region masks as the inputs. The training data for this network are generated from physics simulations seeded from 200 QSM datasets and a set of random background susceptibility distributions. The performance of this neural network was evaluated relative to established background removal methods using 100 in-silico gold standard datasets and clinical susceptibility-weighted imaging datasets. Quantitative and qualitative assessment of the network performance demonstrated the benefits of the trained neural network.