Background field removal method of quantitative susceptibility mapping using deep learning

Inhalt:
Background field removal is a critical step in quantitative susceptibility imaging (QSM) of MRI. Sophisticated harmonic artifact reduction for phase data (SHARP) method is widely used and computation efficiently to remove background field, but it is subject to serious artifact. To reduce artifacts in SHARP method, a deep learning network based on 3D CNN called SHARPnet was designed for background field removal. The reconstructed images with different truncation thresholds using the SHARP principle and the corresponding reconstructed images using laplacian boundary value (LBV) method are used as input and output to train the network. The truncated k-space division (TKD) method is applied to complete the magnetic susceptibility inversion for obtaining QSM. Quantitative measurements of the proposed method was compared with those of the SHARP and the LBV using the public in vivo dataset. The results showed that the proposed method is able to reconstruct the QSM image with less visual error and better quantitative metrics compared with the SHARP method. So SHARPnet can effectively solves the severe artifact in SHARP method by restoring the signal of the threshold truncation region.