Comparative assessment of linear least-squares, nonlinear least-squares, and Patlak graphical method for regional and local quantitative tracer kinetic modeling in cerebral dynamic 18 F-FDG PET.

Purpose Dynamic F-18-FDG PET allows quantitative estimation of cerebral glucose metabolism both at the regional and local (voxel) level. Although sensitive to noise and highly computationally expensive, nonlinear least-squares (NLS) optimization stands as the reference approach for the estimation of the kinetic model parameters. Nevertheless, faster techniques, including linear least-squares (LLS) and Patlak graphical method, have been proposed to deal with high resolution noisy data, representing a more adaptable solution for routine clinical implementation. Former research investigating the relative performance of the available algorithms lack precise evaluation of kinetic parameter estimates under realistic acquisition conditions. Methods The present study aims at the systematic comparison of the feasibility and pertinence of kinetic modeling of dynamic cerebral F-18-FDG PET using NLS, LLS, and Patlak method, based on numerical simulations and patient data. Numerical simulations were used to study the bias and variance of K-1 and K-i parameters estimation under representative noise levels. Patient data allowed to assess the concordance between the three methods at the regional and voxel scale, and to evaluate the robustness of the estimations with respect to patient head motion. Results and Conclusions Our findings indicate that at the regional level NLS and LLS provide kinetic parameter estimates (K-1 and K-i) with similar bias and variance characteristics (K-1 bias +/- relative standard deviation [RSD] 0.0 +/- 5.1% and 0.1% +/- 4.9% for NLS and LLS respectively, K-i bias +/- RSD 0.1% +/- 4.5% and -0.7% +/- 4.4% for NLS and LLS respectively). NLS estimates appear, however, to be slightly less sensitive to patient motion. At the voxel level, provided that patient motion is negligible or corrected, LLS offers an appealing alternative solution for local K-1 mapping. It yields K1 estimates that are highly correlated, with high correlation with NLS values (Pearson's r = 0.95 on actual data) within computations times less than two orders of magnitude lower. Last, Patlak method appears as the most robust and accurate technique for the estimation of K-i values at the regional and voxel scale, with or without head motion. It provides low bias/low variance K-i quantification (bias +/- RSD -1.5 +/- 9.5% and -4.1 +/- 19.7% for Patlak and NLS respectively) as well as smooth parametric images suitable for visual assessment. (c) 2018 American Association of Physicists in Medicine