Effects of statistical noise on graphic analysis of PET neuroreceptor studies

Because of its computational simplicity, the graphic method introduced by Logan et al. is frequently used to analyze time-activity curves of reversible radiotracers measured in brain regions with PET. The graphic method uses a nonlinear transformation of data to variables that have an asymptotically linear relationship. Compared with compartmental analysis of untransformed data, the graphic method enables derivation of regional distribution volumes that are free from assumptions about the underlying compartmental configuration. In this article, we describe statistical bias associated with this nonlinear transformation method.

Theoretic analysis, Monte Carlo simulation, and statistical analysis of PET data were used to test the graphic method for bias.

Mean zero noise is associated with underestimation of distribution volumes when data are analyzed with graphic analysis, whereas this effect does not occur when the same data are analyzed by nonlinear regression and compartmental analysis. Moreover, this effect depends on the magnitude of the distribution volume, so that the bias is more pronounced in regions with high receptor density than regions with low receptor density or no receptors (region of reference).

These results indicate that conventional kinetic analysis of untransformed data is less sensitive to mean zero noise than is graphic analysis of nonlinearly transformed data.