Simulated annealing in pharmacokinetic modeling of PET neuroreceptor studies: accuracy and precision compared with other optimization algorithms

Fitting of PET pharmacokinetic parameters may suffer from bias or unrealistic outcomes, especially for noisy data. There are many readily available optimization algorithms, but each has different characteristics when fitting PET pharmacokinetic models. The purpose of this study is to evaluate the performance of four different types of optimization algorithms, including a modified simulated annealing method, in terms of precision and accuracy of PET pharmacokinetic parameters. The simulated annealing algorithm (SA), called Basin-hoping, was modified for present application. Input data, taken from [¹¹C]-PK11195 neuroreceptor ligand studies, was used to simulate time activity curves at various noise levels. Also the influence of incorrect weighting factors on algorithm performance was studied. Surprisingly, effects of using incorrect but reasonable weighting factors on bias and precision were negligible. Except when extreme and unrealistic weighting factors were used, an increase in bias and decrease in precision was observed. In general the modified SA provided smallest weighted squared residual error and was able to find the global minimum without the need for a proper start parameter selection. However, occasionally better fits were found with the interior-reflective Newton method, but only when implemented using a range of start parameters, centered on the expected value.