Dictionary-Based Compressed Sensing MRI

Magnetic resonance imaging (MRI) is a noninvasive and non-ionizing imaging technique, which is widely utilized in diagnosis due to its excellent ability to visualize both an anatomical structure and a physiological function. However, the main limitation of MRI is its relatively slow imaging modality since the used data, samples of the spatial Fourier transform of the object, are acquired sequentially in time. A straightforward, hardware-based acceleration of MRI can be achieved by partial data acquisition. However, an accurate object reconstruction from reduced amount of measurements still remains a challenge. Actively developing compressed sensing techniques demonstrate that high-quality reconstruction of MR images can be obtained even from fewer Fourier spectral measurements. In this chapter, we describe a method of the compressed sensing MRI reconstruction with the effective split Bregman initialization. The proposed algorithm results in a combination of fast convergence by means of the l1-regularized initialization and high performance by the use of precomputed dictionaries. It demonstrates excellent results on experimental clinic data. In addition, the proposed algorithm is implemented in C++ and optimized for GPU to make the reconstruction time negligible with respect to the data acquisition time.