OBJECTIVE: To reduce variance of the total volume of distribution (V (T)) image, we improved likelihood estimation in graphical analysis (LEGA) for dynamic positron emission tomography (PET) images using maximum a posteriori (MAP). METHODS: In our proposed MAP estimation in graphical analysis (MEGA), a set of time-activity curves (TACs) was formed with V (T) varying in physiological range as a template, and then the most similar TAC was sought out for a given measured TAC in a feature space. In simulation, MEGA was compared with other three methods, Logan graphical analysis (GA), multilinear analysis (MA1), and LEGA using 500 noisy TACs, under each of seven physiological conditions (from 9.9 to 61.5 of V (T)). PET studies of [(11)C]SA4503 were performed in three healthy volunteers. In clinical studies, the V (T) images estimated from MEGA were compared with region of interest (ROI) estimates from a nonlinear least square (NLS) fitting over four brain regions. RESULTS: In the simulation study, the estimated V (T) by GA had a large underestimation (y = 0.27x + 8.72, r (2) = 0.87). Applying the other methods (MA1, LEGA, and MEGA), these noise-induced biases were improved (y = 0.80x + 4.04, r (2) = 0.98; y = 0.85x + 3.05, r (2) = 0.99; y = 0.96x + 1.21, r (2) = 0.99, respectively). MA1 and LEGA produced increased variance of the estimated V (T) in clinical studies. However, MEGA improved signal-to-noise ratio (SNR) in V (T) images with linear correlations between ROI estimates with NLS (y = 0.87x + 5.1, r (2) = 0.96). CONCLUSIONS: MEGA was validated as an alternative strategy of LEGA to improve estimates of V (T) in clinical PET imaging.
OBJECTIVE: To reduce variance of the total volume of distribution (V (T)) image, we improved likelihood estimation in graphical analysis (LEGA) for dynamic positron emission tomography (PET) images using maximum a posteriori (MAP). METHODS: In our proposed MAP estimation in graphical analysis (MEGA), a set of time-activity curves (TACs) was formed with V (T) varying in physiological range as a template, and then the most similar TAC was sought out for a given measured TAC in a feature space. In simulation, MEGA was compared with other three methods, Logan graphical analysis (GA), multilinear analysis (MA1), and LEGA using 500 noisy TACs, under each of seven physiological conditions (from 9.9 to 61.5 of V (T)). PET studies of [(11)C]SA4503 were performed in three healthy volunteers. In clinical studies, the V (T) images estimated from MEGA were compared with region of interest (ROI) estimates from a nonlinear least square (NLS) fitting over four brain regions. RESULTS: In the simulation study, the estimated V (T) by GA had a large underestimation (y = 0.27x + 8.72, r (2) = 0.87). Applying the other methods (MA1, LEGA, and MEGA), these noise-induced biases were improved (y = 0.80x + 4.04, r (2) = 0.98; y = 0.85x + 3.05, r (2) = 0.99; y = 0.96x + 1.21, r (2) = 0.99, respectively). MA1 and LEGA produced increased variance of the estimated V (T) in clinical studies. However, MEGA improved signal-to-noise ratio (SNR) in V (T) images with linear correlations between ROI estimates with NLS (y = 0.87x + 5.1, r (2) = 0.96). CONCLUSIONS: MEGA was validated as an alternative strategy of LEGA to improve estimates of V (T) in clinical PET imaging.