Acceleration in multiple acquisition MRI using deep learning

Abstract

Multiple information can be acquired for the same anatomical structures in Magnetic Resonance Imaging (MRI). Such protocol is termed multiple acquisition, and numerous types of multiple acquisition are possible for specialized purposes. Slice encoding for metal artifact correction (SEMAC) acquires multiple information along z-phase direction, which is later combined for the suppression of metal artifacts along through-plane direction. In routine clinical scans, multi-contrast imaging acquires different information of same anatomy including FLAIR, proton density-, T2-, and T1-weighted contrasts. In multiple phase-cycling balanced steady-state free precession (bSSFP), bSSFP datasets are acquired at multiple phase cycling angles and combined to reconstruct banding artifact-free images. Multiple acquisition in MRI is commonly done, but such acquisition schemes increase scan time. Shortening the scan time, known as the acceleration, has been a widely acknowledged field of study in MRI. Recent development of deep learning has opened new chapters to great achievements in MRI acceleration. This dissertation focuses on acceleration of multiple acquisition MRI protocols mentioned above by using deep learning. Specifically, acceleration of SEMAC was done by using data with lowered amount of z-phase encoding steps, SEMAC factor, to produce high SEMAC factor data. The technique could effectively lower the scan time of SEMAC. Acceleration of multi-contrast and bSSFP data was done through undersampling strategy called varying dimension to produce improved reconstruction quality. The results revealed that varying the undersampled dimension of multiple acquisition data is advantageous over maintaining the dimension. Finally, simultaneous optimization of sampling pattern as well as reconstruction through deep learning is proposed for improving acceleration of multi-contrast imaging. The results showed the advantage of simultaneously optimizing the sampling pattern for joint reconstruction of different contrasts over single contrast optimization and multi-contrast scheme without sampling optimization. Specific sampling behaviors were also investigated for the multi-contrast vs single contrast for sampling optimization.