Performance evaluation of kinetic parameter estimation methods in dynamic FDG-PET studies

Abstract

To evaluate several popular parameter estimation methods for determining the cerebral metabolic rate for glucose and individual kinetic rate constant parameters in 2-deoxy-2-[18F]fluoro-D-glucose positron emission tomography studies. These methods can be divided into two categories: nonlinear estimation methods and linear estimation methods. The nonlinear estimation methods include nonlinear least squares (NLLS), weighted NLLS using noisy tissue time-activity data (WNLLS-N), weighted NLLS using noise-free tissue time-activity data (WNLLS-NF), iteratively reweighted NNLS (IRWNLLS) and nonlinear ridge regression (NLRR) method, whereas the linear estimation methods include Patlak-Gjedde graphical analysis (PGA), linear least squares (LLS), generalized LLS (GLLS), total least squares (TLS) and the basis functions (BF) method. Simulation studies are presented. There are several findings: (i) when the noise level is low, GLLS performs well. However, it exhibits large bias and poor precision especially in k*3 and k*4 when the noise level is high. (ii) BF is a promising method with superior bias and precision properties, and is less affected by the scan duration used. (iii) The weighting factors in the nonlinear estimation methods are important: a good choice of weights can help to make the estimates more accurate and reliable. Weighting based on noisy data should be avoided. (iv) It confirms that PGA is little affected by noise, but the assumptions of PGA could induce bias. It also confirms that 60 min is not long enough to give reliable estimates of k*4 especially for the linear estimation methods LLS, TLS, and GLLS.