HFP-QSMGAN: QSM from homodyne-filtered phase images

Abstract

Purpose: Homodyne filtering is a standard preprocessing step in the estimation of SWI. Unfortunately, SWI is not quantitative, and QSM cannot be accurately estimated from filtered phase images. Compared with gradient-echo sequences suitable for computing QSM, SWI is more readily available and is often the only susceptibility-sensitive sequence acquired in the clinical setting. In this project, we aimed to quantify susceptibility from the homodyne-filtered phase (HFP), acquired for computing susceptibility-weighted images, using convolutional neural networks to solve the compounded problem of (1) computing the solution to the inverse dipole problem, and (2) compensating for the effects of the homodyne filtering. Methods: Two convolutional neural networks, the U-Net and a modified QSMGAN architecture (HFP-QSMGAN), were trained to predict QSM maps at different TEs from HFP images. The QSM maps were quantified from a gradient-echo sequence acquired in the same individuals using total generalized variation (TGV)-QSM. The QSM maps estimated directly from the HFP were also included for comparison. Voxel-wise predictions and, importantly, regional predictions of susceptibility with adjustment to a reference region, were compared. Results: Our results indicate that the U-Net model provides more accurate voxel-wise predictions of susceptibility compared with HFP-QSMGAN and HFP-QSM. However, regional estimates of susceptibility predicted by HFP-QSMGAN are more strongly correlated with the values from TGV–QSM compared with those of U-Net and HFP-QSM. Conclusion: Accurate prediction of susceptibility can be achieved from filtered SWI phase using convolutional neural networks.