Enhanced dynamic cardiac PET imaging using complementary reconstruction

Objectives The noninvasive quantification of myocardial blood flow (MBF) provides assistance in the diagnosis of coronary artery stenosis. Dynamic cardiac PET imaging followed by compartmental modeling allows the measurement of the MBF, showing considerable clinical potentials. However, the conventional approach for reconstructing short time frames of dynamic cardiac PET data leads to poor quantification of tracer uptake due to the limited counts in each frame. The goal of this study is to develop a complementary reconstruction framework for cardiac PET imaging to improve the reconstruction of short time frames and thus to enhance quantitative flow estimation.

Methods To reconstruct the image of a count-limited short time frame, we used a complementary reconstruction algorithm [1] to incorporate the reconstructed image of its complementary frame into the reconstruction of the short time frame. Specifically, the data of the complementary frame with entire acquisition time excluding the short time frame was first reconstructed. The reconstructed image was then forward projected to obtain its contribution in the data space. Then a second reconstruction of the short time frame was executed with data from the entire acquisition and the previous contribution in the data space. For comparison, the conventional way of the short time frame reconstruction including the maximum likelihood (ML) algorithm and the smoothing prior penalized maximum a posteriori (MAP) algorithm were also applied. In order to evaluate the performance of the complementary reconstruction algorithm and to compare it with the ML and the MAP algorithms, we simulated a realistic dynamic cardiac PET dataset with 10 noise realizations. The Rb-82 PET patient organ time activity curves (TACs) including the blood pool and the myocardium were acquired and fitted to generate a set of kinetic parameters. The corresponding parametric images created from the XCAT phantom served as the truth for simulation. Image frames created from the parametric images and the input function were projected to generate the sinogram frames. After the reconstruction of each short time frame, post-filtering was performed. The input function and the myocardial TACs were extracted and corrected for decay. A one-tissue compartment model describing the kinetics of the Rb-82 tracer was applied to analyze the TACs. On the left ventricular (LV) myocardium, we computed voxel-wise estimates of the K1 value. The images of the short time frames reconstructed using the complementary reconstruction algorithm, the ML algorithm, and the MAP algorithm were compared by evaluating the ensemble noise versus bias tradeoff on the LV myocardium. We also compared the estimated K1 values from the aforementioned three algorithms using the tradeoff between the ensemble noise and bias on the LV myocardium.

Results For the reconstruction of short time frames in dynamic cardiac PET, the MAP algorithm reduced the noise at the cost of introducing bias compared to the ML algorithm. The proposed complementary reconstruction algorithm significantly reduced the noise while reducing the bias compared to the MAP algorithm and the ML algorithm. For K1 estimation, the MAP algorithm performed similar to the ML algorithm. The proposed complementary reconstruction algorithm demonstrated improved noise versus bias tradeoff, with a noise reduction of 20% compared to both the ML and the MAP algorithms.

Conclusions For the dynamic cardiac Rb-82 PET imaging, a complementary reconstruction algorithm using data from longer acquisition time to reconstruct the images of the short time frames was implemented. Applying the kinetic modeling in the reconstructed frames, the K1 values were estimated on the LV myocardium. With realistic simulation, we have demonstrated improved performance of the complementary reconstruction algorithm compared to the ML algorithm and the MAP algorithm in both the short time frame reconstruction and the K1 value estimation.