Voxelwise quantification of [(11)C](R)-rolipram PET data: a comparison between model-based and data-driven methods.

This study compared model-based and data-driven methods to assess the best methodology for generating precise and accurate parametric maps of the parameters of interest in [(11)C](R)-rolipram brain positron-emission tomography studies. Parametric images were generated using (1) a two-tissue compartmental model (2TCM) solved with the hierarchical basis function method (H-BFM) linear estimator; (2) data-driven spectral-based methods: standard spectral analysis (std SA) and rank-shaping SA (RS); and (3) the Logan graphical plot. Nonphysiologic VT estimates were eliminated and the remaining ones were compared with the reference values, i.e., those obtained with a voxelwise 2TCM solved with a nonlinear estimator. With regard to voxelwise VT estimates, H-BFM showed the best agreement with weighted nonlinear least square (WNLLS) values and the lowest percentage of mean relative difference (1±1%). All methods showed comparable variability in the relative differences. H-BFM provided the best correlation with WNLLS (y=1.034x-0.013; R(2)=0.973). Despite a slight bias, the other three methods also showed good agreement and high correlation (R(2)>0.96). H-BFM yielded the most reliable voxelwise quantification of [(11)C](R)-rolipram as well as the complete description of the tracer kinetic. The Logan plot represents a valid alternative if only VT estimation is required. Its marginally higher bias was outweighed by a low computational time, ease of implementation, and robustness.