Research Projects

Development of Rapid Compressed Sensing MRI Methods

Michael Lustig, Kurt Keutzer and Mark Murphy

GE Healthcare and UC Discovery 193037

This work aims for developing clinically viable, rapid compressed-sensing Magnetic Resonance Imaging methods. MRI is an excellent tool for diagnosis and monitoring of disease, offering superb soft tissue contrast and high anatomic resolution; unlike computed tomography (CT), particularly attractive is the lack of ionizing radiation. However MRI suffers from several shortcomings, one of which is the inherently slow data acquisition. This has limited the penetration of MRI to applications that require sharp images of fast moving small body parts, such as angiography, cardiovascular imaging, imaging small children and fetal imaging. This limit has led many researchers to look for methods to reduce the amount of acquired data without degrading the image quality. These reduced sampling methods are based on the fact that MRI data is redundant, so the underlying information may be extracted from less data than traditionally considered necessary. Two promising approaches for reduced sampling are sensitivity-encoded imaging (parallel imaging) and compressed sensing. The former uses multiple receivers to accelerate the acquisition, whereas the latter exploits the natural compressibility of medical images. Recently, we have demonstrated that a combined approach of compressed sensing and sensitivity-encoded imaging can a achieve superior image quality at high acceleration factors. Compressed sensing sensitivity-encoded MRI can accelerate the data acquisition, however, the reconstruction involves iterative algorithms that require massive computation. Current reconstruction times are in the order of tens of minutes to several hours, make this technology impractical for routine clinical use. However, recent class of processors; general purpose Graphic Processors (GPGPU) and multi-core CPU architectures have shown great potential for accelerating computations in many application areas, including in MRI reconstruction. Translation of algorithms to these platforms requires high expertise and redesign of the algorithms for parallelization. Our group has great expertise in MRI imaging, compressed sensing algorithms, and translation of algorithms to parallel processing platforms. The aim of this proposal is to help the healthcare industry adopt compressed sensing MRI technology and achieve reconstruction times in the order of seconds to several minutes without compromising the image quality, making compressed sensing MRI a clinically viable option. Our approach is i) To develop new improved efficiency compressed sensing reconstruction algorithms. ii) Translation of the reconstruction algorithms to parallel processing platforms.

Figure 1
Figure 1: Benchmark of reconstruction runtimes of a pro- totype single-core CPU, 12-core CPUs and a GPGPU im- plementation of l1-SPIRiT for several typical 3D matrix sizes and different coil arrays. The reconstruction times are approaching to be clinically useful. The 12-core cpu is approximately equivalent to a single GPGPU and is about 10 times faster than a single CPU.

Figure 2
Figure 2: Our Multi-GPU system at Lucile Packard Children Hospital

Figure 3
Figure 3: Example of image quality: Submilimeter resolution, 8-fold accelerated acquisition of a first pass contrast MR angiography with CS of a 6 years old patient. Pediatric patients have smaller vessels and faster circulation than adults and require much faster imaging. (a) Volume rendering (b) Maximum intensity projection (MIP) and (c) Zoomed MIP showing extraordinary level of details. The data was acquired within 16 seconds compared to 2 min that are required for Nyquist sampling and was reconstructed with our parallel implementation in less than 2min. At that temporal resolution there is no venus contamination in the image.

Mark Murphy, Kurt Keutzer, Shreyas Vasanawala, and Michael Lustig, “Clinically feasible reconstruction time for l1-spirit parallel imaging and compressed sensing MRI,” in Proceedings of the 18th Annual Meeting of ISMRM, Stockhold, Sweeden, 2010, p. 4854.