Comparison of kinetic analysis methods to compute DVR parametric images from PiB-PET dynamic data in regard to noise-induced bias.

Objectives Noise-induced bias on the estimation of distribution volume ratio (DVR) with four kinetic analysis methods was examined in view of parametric image (PI) generation from PiB-PET data of human subjects: (i) relative-equilibrium graphical plot (REplot) (Zhou, NeuroImage 44:661, 2009); (ii) receptor parametric mapping with fixed efflux rate constant (RPM2) (Gunn, NeuroImage 6:279, 1997) which is a modification of SRTM; (iii) Logan graphical analysis with reference region (LGAR); and (iv) two-parameter multilinear reference tissue model (MRTM2) (Ichise, JCBFM 23:1096, 2003).

Methods Two normal controls and two Alzheimer patients underwent a brain dynamic scan for 70-90 min starting injection of ¹¹C-PiB (average 578.4MBq). Regions of interest (ROIs) were placed over the cerebral cortical areas as target regions and cerebellum as reference region. The DVR values for ROIs were estimated by the four methods. Meanwhile, pixel-wise time activity curves (TACs) were generated from the ¹¹C-PiB dynamic data of one Alzheimer patient and fitted with the full kinetic model (MRTM2) to obtain the ideal original TACs. Zero mean Gaussian noise (85% noise corresponding to pixel TAC noise) was added to the original TACs, from which DVR-PIs were generated with the four methods.

Results In the ROI analysis, RPM2 and MRTM2 tended to provide higher DVR estimates than LGAR and REplot. Under noise free condition, REplot generated DVR values slightly lower than the other methods. Under 85% noise, LGAR showed substantial underestimation of DVR, and MRTM2 showed overestimation together with noise expansion, especially in high DVR pixels. On the other hand, both RPM2 and REplot derived much less biased estimation of DVR. The bias due to 85% noise was -0.292+/-0.233 (mean+/-SD) by REplot, -0.113+/-0.161 by RPM2, -1.071+/-0.434 by LGAR, and 0.212+/-0.869 by MRTM2.

Conclusions Among the four methods, RPM2 may derive the most precise and the least biased estimate of DVR.