Unsupervised resolution-agnostic quantitative susceptibility mapping using adaptive instance normalization.

Abstract

Quantitative susceptibility mapping (QSM) is a useful magnetic resonance imaging (MRI) technique that provides the spatial distribution of magnetic susceptibility values of tissues. QSMs can be obtained by deconvolving the dipole kernel from phase images, but the spectral nulls in the dipole kernel make the inversion ill-posed. In recent years, deep learning approaches have shown a comparable QSM reconstruction performance to the classic approaches, in addition to the fast reconstruction time. Most of the existing deep learning methods are, however, based on supervised learning, so matched pairs of input phase images and ground-truth maps are needed. Moreover, it was reported that the deep learning-based methods fail to reconstruct QSM when the resolution of test data is different from the trained resolution. To address this, here we propose an unsupervised resolution-agnostic QSM deep learning method. The proposed method does not require QSM labels for training and reconstructs QSM with various resolutions by using adaptive instance normalization. Experimental results and clinical validation confirm that the proposed method provides accurate QSM with various resolutions compared to other deep learning approaches, and shows competitive performance to the best classical approaches in addition to the ultra-fast reconstruction.