Direct reconstruction (DR) of parametric images from dynamic PET data has been shown to provide substantial noise reduction compared to the conventional indirect reconstruction (IR) approach where frames are first reconstructed and then voxel time-activity curves are fitted to a kinetic model. The main goal was to compare DR and IR, on both within-subject and between-subject variability. Approach. This work evaluated the PMOLAR-1T DR method, using multiple scans of Parkinson's disease (PD) patients with 11CUCB-J, a radioligand for synaptic vesicle glycoprotein 2A (SV2A), a marker for synaptic density. This was achieved by comparing K1, k2, and VT parametric images estimated, at full- and lower-count levels (20%, 10%, and 5%), between DR and IR. Main results. DR delivered considerable improvement, compared to IR, by substantially reducing variability for both within-subject and between-subject analyses, and dramatically reducing noise-induced bias for K1 and VT. Conversely, IR increased the within-subject variability for K1 by 79-353% and for VT by 62-79% across the lower count levels (averaged over regions at matched iterations). The between-subject variability was also increased with IR over DR with an increase of 20-221% for K1 and 45-48% for VT. Even at the full-count level, the between-subject variability was slightly increased for K1 by 4%, but by 24% for VT. Furthermore, at 5% count level, DR delivered comparable variability to IR at 20% counts. The noise-induced %bias, relative to the full-count level, for IR was 3-28% (from 20% to 5% count levels respectively) for K1 and 12 31% for VT, whilst for DR the %bias was only 1% for K1 across count levels, and 7-18% for VT. Significance. To the best of our knowledge, this is the first demonstration that direct-4D reconstruction delivers lower variability and bias not only for within-subject analysis, but also for between-subject analysis.
Gravel et al. (Fri,) studied this question.