Recovering SWI-filtered phase data using deep learning

Abstract

Purpose: To develop a deep neural network to recover filtered phase from clinical MR phase images to enable the computation of QSMs.

Methods: Eighteen deep learning networks were trained to recover combinations of 13 SWI phase-filtering pipelines. SWI-filtered data were computed offline from five multiorientation, multiecho MRI scans yielding 132 3D volumes (118/7/7 training/validation/testing). Two experiments were conducted to show the efficacy of the networks. First, using QSM processing, local fields were computed from the raw phase and subsequently filtered using the SWI-filtering pipelines. The networks were then trained to invert the filtering operation. Second, the trained networks were fine-tuned to recover unfiltered local fields from filtered local fields computed by applying QSM processing to the SWI-filtered phase. Susceptibility maps were computed from the recovered fields and compared with gold standard multiple orientation sampling reconstructions.

Results: Susceptibility maps computed from the raw phase using standard QSM processing have a normalized root mean square error (NRMSE) of 0.732 ± 0.095. Susceptibility maps computed from the recovered phase obtained NRMSEs of 0.725 ± 0.095. The network trained using all 13 processing methods generalized well, obtaining NRMSEs of 0.725 ± 0.89 on filters it has not seen, while matching the reconstruction accuracy of networks trained to recover a single filter.

Conclusion: It is feasible to recover SWI-filtered phase using deep learning. QSM can be computed from the recovered phase from SWI acquisition with comparable accuracy to standard QSM processing.

Keywords: QSM; SWI; deep learning; homodyne; magnetic susceptibility; recovering phase.