Scalable parallelization for high performance tomography

Project Description

In recent years Model-based iterative reconstruction (MBIR) has been under investigation in a number of applications of computed tomography, including medical systems, baggage screening using transmission tomography and X-ray diffraction, materials science using electron microscopy and synchrotron imaging, and more. The nature of MBIR is to incorporate detailed models of the imaging system (geometry, noise characteristics, etc.), and a regularizing prior model of the image volume, directly into the reconstruction process rather than compensate for these characteristics in pre- and post-processing. This allows an advantage in image quality and resolution over direct reconstruction methods that place strong constraints on the assumed system geometry. This advantage has been demonstrated in several published studies in each of the applications listed above.The principal drawback of MBIR is the high computational costs compared to non-iterative methods. This has limited MBIR’s application in fields particularly in which high throughput is a key factor (high-volume security screening), and in which the problem size is exceptionally large (scientific applications like synchrotron imaging and electron microscopy).Our team has invented techniques to better-map the tomographic optimization problem onto hardware in a way that dramatically improves memory coalescence and cache hit rates, as well as exploits hierarchical levels of parallelism.These techniques have produced speedups of 2 to 3 orders of magnitude. These improvements have the potential to mainstream the application of MBIR and expand the capability of a wide range of imaging systems. Further, experiments on single-unit GPU have demonstrated 4.4x speedup over a state-of-the-art implementation on a 16-core iso-power CPU, which point to GPU as the most cost-effective platform for this problem.