CT Image Reconstruction on Intel Larrabee using Half Precision Floating-Point Values

摘要

Computed tomography is a highly quantitative modality where the data values processed are of a high dynamic range and therefore computing CT images requires high accuracy and high precision during all processing steps. Nevertheless, memory bandwidth is one of the major bottlenecks in today's reconstruction algorithms and it would speed up the reconstruction process to represent a datum with only 16 bit instead of 32 bit. Intel's new Larrabee x86 architecture offers a promising processing of the half precision floating-point value that has the potential to be useful for CT image reconstruction. This study evaluates the image quality of CT image reconstruction based on such half values and compares the results to the standard case of using float values. Our findings are that it is possible to apply half precision data for rawdata and reconstructed volume if some conditions hold.