Implementation and Evaluation of Uncertainty Estimation for Advanced Pharmacokinetic Models in DCE-MRI

Abstract: The goal of this report is to develop code in Python that implements and evaluates a multivariate linear error propagation for commonly used DCE-MRI models. Two pharmoacokinetic models (PK-models) were tested; modified Kety model and two-compartment exchange model. The analysis compared a Monte carlo simulated signal and a simple analytic model for uncertainty. The purpose of comparing was to have the Monte carlo method and the linear error estimation method reasonably close in terms of estimation of the error of the PK-parameters (parameters are specific for the PK-model). The analysis included an inspection of the resulting plots of the Coefficent of Variation (CV) of the estimation of the PK-parameters, when increasing the ratio of the true value and the added error for each of the input parameters. Generally the CV increased, I.e. the precision decreased, as the noise level increased. For both PK-models, error in injection time of contrast agent were the input parameter with the least tolerance of noise. By comparing the results from both methods, the conclusion is to implement the modified Kety model as the PK-model used in MICE Toolkit (NONPI Medical AB, Umeå, Sweden) and keep the noise to signal level below 15\% in order to obtain precision of the results to less than 33\% error. We have verified the uncertainty estimation method works well for the Modified Kety model but not so well for the two-compartment exchange model. The scope of this study only included two PK-models and one signal model, therefore it would be beneficial to test the linear estimation method used here on other PK-models and signal models to find what models it works well for.