Image Reconstruction and Motion Compensation Methods for Fast MRI Chaoping

Research output: Types of ThesisDoctoral ThesisInternal


Magnetic resonance imaging (MRI) is one of the most widely used methods in medical imaging and can provide various soft tissue contrasts for various anatomies. MRI forms images from scanner acquired k-space signals using various imaging pulse sequences. Such acquisitions can take long times, which is one of the major drawbacks of MRI. To speed up, people have developed numerous fast imaging sequences, trying to acquire more k-space signals in shorter time. Another way to accelerate the imaging is subsampling the k-space, which imposes challenges for reconstruction. The clinical MRI scans are usually performed with a multi-channel receive coil and/or using multiple pulse sequence settings. The signal correlation among the channels or different sequence settings provides opportunities for reconstructing subsampled k-space. This thesis proposes multiple auto-calibrated reconstruction methods by exploiting such signal correlation. In this thesis, the autocalibration is not only used to accelerate the imaging but also to compensate for the motion that happens during the scanning. In addition to the conventional linear way of exploiting the signal correlation, we also explored non-linear autocalibration using a neural network. Finally, we explored the potential of acceleration with a deep learning inverse problem solver by exploiting further the shared knowledge between image reconstruction and relaxometry parameter mapping for quantitative MRI.

In Chapter 2 we propose a reconstruction method, APIR4GRASE, for the GRASE sequence. The GRASE sequence interleaves the spin echoes and gradient echoes in the acquisition and therefore suffers from modulation artifacts from the mixed echo types. The APIR4GRASE integrates autocalibrated parallel imaging reconstruction among the different echo types with the selected optimized sampling patterns, achieving better image quality with less aliasing artifacts and noise amplification than the conventional 3D-GRASE. It reconstructs images using all echo types as virtual coil channels, in contrast to GRAPPA which individually reconstructs each echo type. The optimal sampling patterns of k-space associated with the proposed reconstruction method require exhaustive search, although for similar anatomies and scanning settings, a one time search with retrospective subsampling would be near optimal for prospective acquisitions. APIR4GRASE assumes spatially smooth T2* decay between the spin echo and the gradient echo, which is typically true in the brain. In a prospective acquisition, it achieved 0.8 mm 3D isotropic T2-weighted brain imaging with scan time of 5.5 minutes, whereas the alternative conventional GRASE SORT imaging with a subsampling factor of 2 took 9.4 min.

In Chapter 3 we propose APIR4EMC which reconstructs multi-contrast images with autocalibrated parallel imaging reconstruction by adding contrasts as virtual coils. It is extended from APIR4GRASE (Chapter 2) and reconstructs multiple contrasts instead of different echo types in a sequence. In the multi-contrast imaging, different contrasts are acquired separately with different protocols, and the signal evolution along the echo train is therefore also different. We compensate for the difference with stabilization and Fermi filtering, which has been proven in the experiments to be able to improve the image quality. We jointly optimized the k-space sampling patterns of the multi-contrast acquisitions with exhaustive search, similar as in Chapter 1. With APIR4EMC, we improve the image quality over GRAPPA, and achieved 1 mm 3D isotropic in-vivo multi-contrast (T1, T1-Fatsat, T2, PD, FLAIR) brain imaging with scan time of 7.5 minutes.

In Chapter 4 we propose a retrospective translational motion compensation method for parallel imaging 3D FSE acquisitions. Assuming no motion within each echo train of the FSE, we estimate the motion parameters of every echo train. The method relies on the optimization of data consistency in the fully sampled ACS region. To allow this, the ACS region is expected to contain echoes from every train, for which we propose a radial spokes view ordering for the 3D Cartesian k-space. The optimization is solved by alternating the estimation between the GRAPPA prediction kernel and motion parameters. Experiments with simulated motion and acquired motion in in-vivo acquisitions results show that the proposed method is able to substantially reduce the motion artifacts of the motion corrupted acquisitions.

In Chapter 5 we propose a scan specific, auto-calibrated k-space completion method for parallel imaging, APIR-Net, to reconstruct the full k-space from an undersampled k-space by exploiting the redundancy among the multiple channels in the receive coil. The proposed APIR-Net is featured with a decreasing number of feature maps when encoding layer goes deeper, and a constant spatial size for all feature maps. Unlike the conventional parallel imaging methods that estimate the prediction kernel and perform interpolation in a linear way, APIR-Net is able to learn nonlinear relations among sampled and unsampled positions in k-space. The experiments show that APIR-Net was able to reduce noise amplification and increase the visual image quality compared to the state-of-the-art ESPIRiT and RAKI methods in both phantom and in-vivo experiments, making APIR-Net a promising alternative in low SNR acquisitions.

In Chapter 6 we propose qRIM to accelerate the quantitative MR imaging. It embeds a unified forward model for joint reconstruction and R2*-mapping from sparse data in a Recurrent Inference Machine (RIM), an iterative inverse problem solving network. The integrated prior of the unified forward model facilitates the exploitation of the shared knowledge between the reconstruction and parameter estimation, including the redundancy among TEs. In the experiments, the proposed qRIM reduced the mapping error as well as imaging blurriness compared to the alternative sequential model of image reconstruction and parameter fitting, and the reduction of the reconstruction error increased with acceleration factor. With qRIM, we achieved a stable R2* mapping of the human subcortex up to 9-fold acceleration.

Finally, in Chapter 7, we discussed the contributions and limitations of this thesis, and proposed a few future perspectives.

In conclusion, this thesis presented several new autocalibration methods to improve image quality of reconstruction in acquisitions with multiple imaging sequence settings. We also presented a neural network that exploits non-linear autocalibration and which is able to reconstruct better image for low SNR acquisitions than the state-of-the-art ESPIRiT and RAKI methods. In addition to reconstruction, we also presented a novel retrospective translational motion compensation method by exploiting autocalibrated signals with a specifically designed view ordering for the parallel imaging 3D FSE acquisitions. Further exploiting the shared knowledge between the image reconstruction and parameter mapping, we presented the qRIM method that is able to improve the reconstruction quality of R2*. The image reconstruction and motion compensation methods proposed in this thesis may contribute to the implementation of faster MRI methods in clinical practice.
Original languageEnglish
Awarding Institution
  • Erasmus University Rotterdam
  • Klein, Stefan, Supervisor
  • Poot, Dirk, Co-supervisor
Award date23 Nov 2022
Place of PublicationRotterdam
Print ISBNs978-94-6469-070-5
Publication statusPublished - 23 Nov 2022


Dive into the research topics of 'Image Reconstruction and Motion Compensation Methods for Fast MRI Chaoping'. Together they form a unique fingerprint.

Cite this