CT image reconstruction with half precision floating-point values

Purpose: Analytic CT image reconstruction is a computationally demanding task. Currently, the even more demanding iterative reconstruction algorithms find their way into clinical routine because their image quality is superior to analytic image reconstruction. The authors thoroughly analyze a so far unconsidered but valuable tool of tomorrow’s reconstruction hardware (CPU and GPU) that allows implementing the forward projection and backprojection steps, which are the computationally most demanding parts of any reconstruction algorithm, much more efficiently.Methods: Instead of the standard 32 bit floating-point values (float), a recently standardized floating-point value with 16 bit (half) is adopted for data representation in image domain and in rawdata domain. The reduction in the total data amount reduces the traffic on the memory bus, which is the bottleneck of today’s high-performance algorithms, by 50%. In CT simulations and CT measurements, float reconstructions (gold standard) and half reconstructions are visually compared via difference images and by quantitative image quality evaluation. This is done for analytical reconstruction (filtered backprojection) and iterative reconstruction (ordered subset SART).Results: The magnitude of quantization noise, which is caused by a reduction in the data precision of both rawdata and image data during image reconstruction, is negligible. This is clearly shown for filtered backprojection and iterative ordered subset SART reconstruction. In filtered backprojection, the implementation of the backprojection should be optimized for low data precision if the image data are represented in half format. In ordered subset SART image reconstruction, no adaptations are necessary and the convergence speed remains unchanged.Conclusions: Half precision floating-point values allow to speed up CT image reconstruction without compromising image quality.