Component analysis algorithms for fast multi-tracer PET signal-separation

Objectives Rapid multi-tracer PET can image 2-3 tracers in a single scan to characterize multiple aspects of physiology and function. Dynamic imaging with staggered injections is used, and kinetic constraints are applied to recover individual-tracer signals from multi-tracer datasets. The kinetic constraints are often imposed using multi-tracer compartment modeling. While effective, such algorithms are slow, involve multidimensional nonlinear fits, and require knowledge of the arterial input function for each tracer. This work proposes component-analysis based algorithms for rapid multi-tracer PET signal-separation designed to overcome these limitations.

Methods Linear sets of temporal basis functions for each tracers’ allowed kinetic behavior were formed from principal component analysis (PCA) of population time-activity curves combining simulated time-activity curves with measured image components. The resulting linear systems were orthogonal intra-tracer but non-orthogonal between tracers. The multi-tracer signal-separation algorithm first performs linear weighted-least squares fits to these bases for each image voxel. The fitted time-activity curves for each voxel are then used to estimate the proportional contribution of each tracer to each timepoint of the 4D multi-tracer image according to linear mixture analysis. Each voxel of the multi-tracer input image is then separated to form individual-tracer images, preserving the total sum at each point.

Results The PCA-based algorithm was evaluated for several tracer combinations using both simulations and clinically acquired images in patients with primary brain tumors (18F-FDG, 18F-FLT, and 11C-acetate). Single-scan dual-state rest+stress myocardial perfusion images were also tested (13N-ammonia). Signal-separation performance was similar to that for full multi-tracer compartment modeling in a fraction of the time, providing voxelwise image separation in just a few seconds CPU time.

Conclusions The proposed component-analysis algorithm for multi-tracer PET signal-separation provides fast and robust voxel-by-voxel separation of multi-tracer images.

Research Support R01 CA135556