Abstract
622
Objectives Rapid multi-tracer PET images 2-3 tracers in a single dynamic scan. Here, tracer administrations are staggered in time, and kinetic constraints are employed to recover individual-tracer signals from multi-tracer data. The prototypical signal-separation approach is to simultaneously apply and fit compartment models for all tracers. Such multi-tracer modeling results in large, multidimensional nonlinear fits fraught with local minima and complex topology. We recently described Separable Parameter Space techniques for fitting 1-3 tissue compartment models to dynamic imaging data. This work investigates explores the value and performance of Separable Parameter Space fitting for multi-tracer PET data.
Methods The technique effectively separates the linear and nonlinear aspects of the fit, greatly reducing the dimensionality of the nonlinear sub-problem. The new algorithm was applied to clinical research studies of rapid dual tracer PET, including: rest+stress myocardial perfusion PET with [13]N-ammonia; FLT+FDG imaging of lung tumors; and FLT+[11]C-acetate imaging of primary brain tumors. These examples cover dual-tracer modeling examples with 2-tissue compartments and 3-4 rate parameters for each tracer. Iterative and exhaustive search Separable Parameter Space fits were performed and compared against conventional non-negative least-squares fits.
Results Inspection of nonlinear objective functions for the Separate Parameter Space technique demonstrated a marked reduction—and often complete elimination—of local minima for dual-tracer fitting throughout typical parameter ranges. Conventional iterative fits were trapped by local minima in over 50% of cases, whereas the new algorithm converged to the true global minimum for 100% of the cases studied. Fitting times were also reduced by factors of 10-100, providing robust fits in 7.7-48.9ms CPU time.
Conclusions Separable Parameter Space techniques provide very fast and robust multi-tracer fits, overcoming a substantial hurdle in the development of useable algorithms for multi-tracer PET imaging.
Research Support R01CA135556