Using bootstrap identifiability as a metric for model selection for dynamic [/sup 11/C]DASB PET data

Numerous tracer kinetic models have been developed for estimation of neuroreceptor binding parameters from dynamic PET and SPECT brain studies. We have used the bootstrap technique to determine the variability of the parameter estimation as an aid in selecting the most appropriate kinetic model to use. This technique made it possible to take into account different sources of variability. We applied the method to data from 11 healthy subjects, each one scanned twice with the PET serotonin transporter ligand [11C]DASB. Tracer binding was quantified for different brain regions by kinetic analysis, based on metabolite corrected arterial plasma input functions. Six different analysis methods were used, including iterative as well as non-iterative implementations of 1- and 2-tissue compartmental models (1TC, 2TC, 1TCNI, 2TCNI), likelihood estimation in graphical analysis (LEGA), and basis pursuit (Basis). We applied the bootstrap technique to the PET data, as well as to the plasma and metabolite data. Standard errors (SE) were calculated for the total volume distribution (VT), as well as different binding potential estimates. The average and standard deviation (SD) of the estimated SE values were calculated across subjects. For comparison, we also estimated the variability of the outcome measures by bootstrapping only the tissue data. The results of the full bootstrap analysis showed that Basis was in general the best method. However, when only the tissue data were bootstrapped, the results indicated that 1TCNI was best. This shows that it can be important to take into account all sources of variability when using bootstrap identifiability for model selection