MIXTURE MODELING FOR DYNAMIC PET DATA

Standard kinetic modeling of dynamic positron emission tomography (PET) data requires specifying a compartment structure and fitting the appropriate kinetic model using nonlinear least squares algorithms separately for each voxel in the brain. This approach is not completely satisfactory because of a natural reluctance researchers have to specifying a particular compartmental model to be applied to all voxels and, in addition, there are parameter identifiability issues for all but the simplest models. This paper presents new methodology for modeling dynamic PET data that works by "borrowing strength" across all voxels, expressing each voxel's data as a linear combination of a small number of components that are estimated from the data. Though based on a kinetic modeling structure, it does not require a choice of compartmental system and allows for data-adaptive choice of model order. The spatial autocorrelation throughout the brain is modeled with a conditional autoregressive (CAR) model. Estimation of model parameters is accomplished through iterative optimization based on nonlinear weighted least squares, and selection of the number of components is based on a modified information criterion. This methodology may be applied either at a voxel-level or in a region of interest (ROI) analysis. Performance of the method is evaluated with simulated and real data.