Synopsis

Keywords: Image acquisition: Fast imaging, Image acquisition: Sequences, Image acquisition: Reconstruction

Since the advent of MRI technology, image acquisition acceleration has consistently been one of the most prominent research topics in the field. Up to now, sampling acceleration still remains a central focus in MR technology research. Numerous acceleration techniques have been developed, with some already incorporated into products. These techniques include fast imaging sequences, partial Fourier-encoding, parallel imaging, compressed sensing, deep learning, and combinations thereof. Imaging acceleration techniques are widely used in anatomical, functional, and dynamic imaging. This lecture will provide a brief explanation of the fundamental principles behind some of these techniques.

Introduction and Background

Since the advent of MRI technology, image acquisition acceleration has consistently been one of the most prominent research topics in the field. Up to now, sampling acceleration still remains a central focus in MR technology research. Numerous acceleration techniques have been developed, with some already incorporated into products. These techniques include fast imaging sequences, partial Fourier-encoding, parallel imaging, compressed sensing, deep learning, and combinations thereof. Imaging acceleration techniques are widely used in anatomical, functional, and dynamic imaging. This lecture will provide a brief explanation of the fundamental principles behind some of these techniques.

Parallel Imaging

Parallel imaging has emerged as one of the most widely used acceleration techniques in MRI. It exploits the spatial sensitivity information from phase array RF coils to reconstruct images from undersampled k-space data below the Nyquist rate. Since fewer k-space lines are acquired, parallel imaging reduces the overall scan time and mitigates motion artifacts. Two popular parallel imaging methods, SENSE (Sensitivity Encoding)(1) and GRAPPA (Generalized Autocalibrating Partially Parallel Acquisitions)(2) have been used for various clinical and research applications. SENSE unfolds aliased images in image domain and requires precise knowledge of the coil sensitivity profiles which are usually acquired separately. GRAPPA fills in the undersampled k space and requires interpolation kernel information from calibration data, which are obtained from either fully-sampled central k-space regions or from separate low-resolution scans. Due to practical limitations such as noise and imperfect coil geometry, acceleration factors of 2 ~ 3 are commonly used with an 8- or 16-channel RF coil for SENSE or GRPPA. Simultaneous multi-slice (SMS) imaging is also a type of parallel imaging. It acquires multiple slices simultaneously by exciting them using a single RF pulse with different frequency bands. For image reconstruction, parallel imaging processing method such as slice-GRAPPA can be used to separate the aliased slices using coil sensitivity information.

Compressed Sensing

Compressed sensing (CS) (3) is an acceleration technique in MRI that exploits the inherent sparsity or compressibility of MR images in a suitable transform domain (e.g., wavelets or total variation) to reconstruct images from significantly fewer samples than required by the Shannon sampling theorem. For pulse sequence design, CS uses non-uniform and pseudo-random k-space sampling patterns, such as variable-density Poisson-disc. This incoherent sampling can spread aliasing artifacts in the image domain, making them more amenable to removal by the sparsity-promoting reconstruction. For image reconstruction, CS usually uses non-linear reconstruction algorithms to recover the sparse image from the undersampled data. These algorithms often minimize an objective function that consists of a data fidelity term (to ensure consistency with the acquired data) and a regularization term (to promote sparsity in the transform domain). Widely used optimization methods in CS reconstruction include the iterative shrinkage-thresholding algorithm (ISTA), the fast iterative shrinkage-thresholding algorithm (FISTA), and the alternating direction method of multipliers (ADMM).
CS and parallel imaging techniques such as SENSE can be combined to achieve higher acceleration factors while maintaining satisfactory image quality. Then the objective function to be minimized during the reconstruction has an extra term: the parallel imaging term, which incorporates the coil sensitivity information to help remove aliasing artifacts.

Deep learning

Parallel imaging relies on spatial sensitivity information of phase array coils for the reconstruction of undersampled data. CS reconstruction techniques not only utilize the redundant information from multiple channels but also exploits the prior information that MR images have sparsity in certain transform domains. Without introducing additional prior information, CS can achieve the highest acceleration factors of 3~4 for one-dimensional undersampling. For two-dimensional random undersampling, the highest acceleration factor can be around 10. Additionally, the reconstruction time for CS is relatively long.
In 2015, Wang and others first introduced deep learning neural networks into CS reconstruction (4). Deep learning have the characteristic of non-linear modeling and can capture non-linear information in signals. At the same time, neural networks use end-to-end methods, training repeatedly so that their parameters can fit the reconstruction process of MRI images from undersampled to fully sampled. As a result, they achieve better results in obtaining MRI image prior information compared to dictionary learning, further improving reconstruction quality. Since then, MRI image reconstruction techniques based on deep learning have become a mainstream research direction in MRI reconstruction. Compared to traditional CS reconstruction algorithms, deep learning-based MRI image reconstruction has higher quality and faster speed. Thus deep learning has already a significant impact on imaging acceleration, although it is not clear how this technique evolves into products eventually.

Conclusion

Imaging acceleration techniques enable higher sampling efficiency and thus play an important role in MRI applications. By combining the various sampling acceleration techniques, even higher acceleration factors can be achieved, further pushing the speed of MRI acquisitions. With the advancement of technology, we believe that MR imaging will become even faster, providing more reliable information for diagnosis and treatment.

Acknowledgements

No acknowledgement found.