Effect of reconstruction and filtering on kinetic parameter estimation bias and reliability for dynamic SPECT : a simulation study

Abstract

Dynamic SPECT has the potential to provide absolute physiological parameter estimates. However, the low sensitivity of SPECT typically results in very noisy dynamic SPECT data. Filtering can reduce the noise, but at the expense of degrading the already poor resolution further. The effect of reconstruction parameters, post-reconstruction filtering and resolution recovery on kinetic parameter estimation bias and reliability was systematically investigated. Dynamic projection data were generated using Monte Carlo (MC) simulations of a mathematical brain phantom at 10 different levels of Poisson noise. The projection data were reconstructed with OSEM with varying numbers of iterations and subsets and were filtered with three-dimensional (3-D) Gaussian filters with varying FWHM. Bias and reliability of the main parameters of interest (K₁,V[sub d], and binding potential) for thalamus, cerebellum, and frontal cortex were estimated for the three-compartment model fits to the tissue time-activity curves derived from the reconstructed data. Reliability (standard deviation) of parameter estimates was obtained with the Bootstrap MC technique, which showed good agreement with conventional MC in a subset of data sets, but required only a small fraction of conventional MC computation time. Post-reconstruction filtering increased bias, without improving the reliability of parameter estimates and, hence, no post-reconstruction filtering is recommended. For reconstructions without resolution recovery, an effective number of 40 iterations overall provided the best tradeoff between bias and reliability for the range of noise levels studied. Resolution recovery achieved a modest reduction in bias.