With the increasing reliance of doctors on imaging procedures, not only visualization needs to be optimized,
but the reconstruction of the volumes from the scanner output is another bottleneck. Accelerating the computationally
intensive reconstruction process improves the medical work flow, matches the reconstruction speed
to the acquisition speed, and allows fast batch processing and interactive or near-interactive parameter tuning.
Recently, much effort has been focused on using the computational power of graphics processing units (GPUs)
for general purpose computations. This paper presents a GPU-accelerated implementation of single photon
emission computed tomography (SPECT) reconstruction based on an ordered-subset expectation maximization
algorithm. The algorithm uses models for the point-spread-function (PSF) to improve spatial resolution in the
reconstruction images. Instead of computing the PSF directly, it is modeled as efficient blurring of slabs on the
GPU in order to accelerate the process. The algorithm for the calculation of accumulated attenuation factors
that allows correcting the generated volume according to the attenuation properties of the volume is optimized
for processing on the GPU. Since these factors can be reused between different iterations, a cache is used that is
adapted to different sizes of the video memory so that only those factors have to be recomputed that do not fit
onto graphics memory. These improvements make the reconstruction of typical SPECT volume near interactive.