Compressed sensing in dynamic MRI

Urs Gamper, Peter Boesiger, Sebastian Kozerke

Magnetic Resonance in Medicine · 2008 · 564 citations · 28 references

Concepts

TL;DR

Compressive sensing offers promise for MRI, yet typical MR images lack sparsity, which dynamic imaging can provide by exploiting temporal Fourier sparsity and random phase‑encoding undersampling. This study evaluates the feasibility of compressive sensing for accelerating dynamic MRI. The authors compare simulated datasets across reduction factors, noise, and sparsity, and test in vivo cardiac cine and carotid velocity data against k‑t BLAST reconstructions. When data sparsity and SNR are sufficient, CS yields higher temporal fidelity than k‑t BLAST for the examined datasets. Magn Reson Med 59:365–373, 2008 and © 2008 Wiley‑Liss, Inc.

Abstract

Abstract Recent theoretical advances in the field of compressive sampling—also referred to as compressed sensing (CS)—hold considerable promise for practical applications in MRI, but the fundamental condition of sparsity required in the CS framework is usually not fulfilled in MR images. However, in dynamic imaging, data sparsity can readily be introduced by applying the Fourier transformation along the temporal dimension assuming that only parts of the field‐of‐view (FOV) change at a high temporal rate while other parts remain stationary or change slowly. The second condition for CS, random sampling, can easily be realized by randomly skipping phase‐encoding lines in each dynamic frame. In this work, the feasibility of the CS framework for accelerated dynamic MRI is assessed. Simulated datasets are used to compare the reconstruction results for different reduction factors, noise, and sparsity levels. In vivo cardiac cine data and Fourier‐encoded velocity data of the carotid artery are used to test the reconstruction performance relative to k‐t broad‐use linear acquisition speed‐up technique ( k‐t BLAST) reconstructions. Given sufficient data sparsity and base signal‐to‐noise ratio (SNR), CS is demonstrated to result in improved temporal fidelity compared to k‐t BLAST reconstructions for the example data sets used in this work. Magn Reson Med 59:365–373, 2008. © 2008 Wiley‐Liss, Inc.

References

28