We present new implementations of the Maximum Likelihood Expectation Maximization (EM) algorithm and the related Ordered Subset EM (OSEM) algorithm. Our implementation is based on modern graphics hardware and achieves speedups of over eight times current software implementation, while reducing the RAM required to practical amounts for today’s PC’s. This is significant as it will make this algorithm practical for clinical use. In order to achieve a large speed up, we present bit splitting over different color channels as an accumulation strategy. We also present a novel hardware implementation for volume rendering emission data without loss of accuracy. Improved results are achieved through incorporation of attenuation correction with only a small speed penalty.

Top