Keywords: change-point estimation, onset of change model, PET imaging, graphical analysis, simulation
Graphical analysis offers a simplified alternative to kinetic modeling when quantifying Positron Emission Tomography (PET) brain images. Such analysis relies on selecting a time-point t*after which the relationship between the variables involved in the analysis is approximately linear. t* can be determined by visual inspection of the data, but this is subjective and impractical when quantifying many images. Automatic procedures for choosing t* require specification of an arbitrary threshold (e.g., a bound on the relative size of residuals). We propose an alternative fully automatic approach based on how well the graphical model fits the data. For each candidate t* value, we fit all data points for which t>t* according to a likelihood-based procedure for graphical analysis. The optimal t* is then automatically selected based on these residuals as the solution to a problem in change-point estimation, by applying an onset-of-trend change-point model to the estimates of the noise level. We apply this procedure to both simulated and clinical PET human data.