Statistical comparison of dynamic contrast-enhanced MRI pharmacokinetic models in human breast cancer

Abstract

By fitting dynamic contrast-enhanced MRI data to an appropriate pharmacokinetic model, quantitative physiological parameters can be estimated. In this study, we compare four different models by applying four statistical measures to assess their ability to describe dynamic contrast-enhanced MRI data obtained in 28 human breast cancer patient sets: the chi-square test (χ(2)), Durbin-Watson statistic, Akaike information criterion, and Bayesian information criterion. The pharmacokinetic models include the fast exchange limit model with (FXL_v(p)) and without (FXL) a plasma component, and the fast and slow exchange regime models (FXR and SXR, respectively). The results show that the FXL_v(p) and FXR models yielded the smallest χ(2) in 45.64 and 47.53% of the voxels, respectively; they also had the smallest number of voxels showing serial correlation with 0.71 and 2.33%, respectively. The Akaike information criterion indicated that the FXL_v(p) and FXR models were preferred in 42.84 and 46.59% of the voxels, respectively. The Bayesian information criterion also indicated the FXL_v(p) and FXR models were preferred in 39.39 and 45.25% of the voxels, respectively. Thus, these four metrics indicate that the FXL_v(p) and the FXR models provide the most complete statistical description of dynamic contrast-enhanced MRI time courses for the patients selected in this study.