Assessment of kinetic modeling quality of fit by cluster analysis of residuals: Application to direct reconstruction of cardiac PET data

Direct reconstruction algorithms for PET are capable of producing lower variance parametric images than those from the traditional frame-based approach. However, if the selected kinetic model is not appropriate for all voxels in the field-of-view, structured residuals from poor model fits can propagate bias into neighboring regions where the model is accurate. We previously proposed a direct reconstruction algorithm for cardiac PET that fits the kinetic model only at voxels in a pre-specified mask, and uses cubic B-splines elsewhere. Here we present an automated approach to generate this mask, based on k-means clustering of the residuals from a frame-based fit. In this application, the cluster-based mask allowed the kinetic model to be fit to many more voxels than those defined by a manually drawn myocardium region. However, the resulting parameter estimates had comparable agreement with frame-based estimates as those generated using the more restrictive mask. The hybrid method with either the manually defined or cluster-based mask was in better agreement with the frame-based method than a direct reconstruction using the kinetic model in all voxels, suggesting that the hybrid method with a cluster-based mask can reduce bias from error propagation of non-model fits.