High Performance Computing for Dynamic BioMedical Imaging
Award year: 2005-2006
The imaging of dynamic phenomena in the human body presents some of the greatest challenges in biomedical imaging. Important applications of this so-called 4D imaging include cardiac imaging, functional brain imaging and interventional imaging, or image-guided surgery imaging. The main challenge in 4D imaging is to obtain, at the same time, both high resolution in time, and high resolution in 3 dimensional space of internal body structure and function. The particular imaging methods we consider in this project are near-infra-red imaging, x-ray computerized tomography (CT), and magnetic resonance imaging (MRI). These methods have either high temporal but low spatial resolution (infra-red imaging), high spatial but low temporal resolution (CT), or, in the case of MRI, offer a tradeoff between spatial and temporal resolution, but not high values for both.
In all cases, the physics of the imaging process appear to limit the amount of data that can be extracted. The primary approach to overcoming these limitations has been expensive development of the physical acquisition hardware, but this approach too, has its economic and physical limits. The remaining gap between the available and desired spatial or temporal resolutions can be as high as factor of 10.
In this project, we will use computational power provided by NCSA supercomputing platforms to overcome some of the limitations of current dynamic imaging techniques. Computational power will enable the use of new methods for acquisition of the data and for the formation of images, which the PI has developed over the past several years. The methods have three components: (i) adaptive acquisition of the data; (ii) new image formation methods; and (iii) efficient computational algorithms. By tailoring the acquisition to the spatio-temporal characteristics of the particular imaged object, the amount relevant data extracted is maximized. Using this maximally informative data requires new, non-linear image formation methods. Finally, efficient algorithms reduce the computational requirements. In some cases, these methods can provide dramatic, order of magnitude or greater improvements in temporal and/or spatial resolution. However, because of the heavy computational requirements of these methods, they are currently infeasible for real-time 4D imaging.
To accelerate these new dynamic imaging methods, we will study with NCSA researchers the use of high-performance FPGA-accelerated reconfigurable computing platforms such as the Cray XD1. This platform is attractive, because of the large data throughput requirements in 4D biomedical imaging, the intense computation required by the new methods, and the fine-grained parallelism that characterizes these methods.
Determining how well the new methods map to this architecture, and the achievable performance, will ultimately determine whether the new methods become feasible, and whether similar computing platforms can be used in actual diagnostic imaging systems.