To the best of our knowledge, it is the first time that a phantom study is performed to identify the best reconstruction algorithm to compare the LA uptake in patient cardiac FDG studies between an analog and a digital TOF-PET/CT. As the older analog TOF-PET/CT system does not include any RR option in the reconstruction algorithm, a third party post-reconstruction RR based on the Richardson-Lucy method was considered.
While a third party post-reconstruction RR method clearly improves the spatial resolution and cardiac walls activity recoveries by using a constant PSF (see figure 1 and table 1), it is not optimized to the PET image specificities like the manufacturer algorithm that includes a RR that takes into account the variations of the PSF according to the location in the field of view [13]. This results in noisier images, especially in the case of the analog TOF-PET/CT whose original reconstruction is already more impacted by noise (figure 1). Noisier images will result in less reliable activity measurements as very sensitive to the VOIs drawing. It should then be recommended to use the best TOF-PET/CT available to perform this kind of quantitative studies.
Figure 3B shows, for the Vereos TOF-PET/CT, that when no RR or only a light RR (only 1 iteration) is performed, the SNR behaves like an increasing function of the LA activity recovery, while for higher RR iteration numbers, it becomes a decreasing function. This induces that a compromise has to be made between the LA activity recovery and the image reading quality. A post filtering of the image is still possible to make a noisier image more readable. However, one should be careful when pushing the RR algorithms too far (disregarding the SNR) as one might end up with artefactual images that are no more representative of the activity distribution, typically by producing edge artefacts [18,19].
Based on the evaluation of the image quality by our nuclear medicine physicians, we selected as best reconstruction parameters of the OSEM algorithm the values of 4 iterations and 15 subsets as these values correspond to the beginning of the convergence plateau of the activity recovery (see figure 3A), and higher values lead to noisier images. For the manufacturer RR parameters we decided to limit the number of iterations to 5 as the SNR drops quite fast for higher values (see figure 3B). In the present study we did not play with the regularization parameter of the manufacturer RR algorithm that was fixed to the recommended value of 6. Optimizing this parameter, whose function is to drive the image smoothness, might still improve the image quality by reducing the noise level but we do not expect it to alter much the optimized set of parameters found in the present quantitative study. Together these parameters provide a LA activity recovery of about 70%.
The FDG cardiac studies performed on both TOF-PET/CT with the same patient nicely confirmed the results obtained with the phantom. Compared to the phantom the patient heart was about 10 percent larger, but ventricular and atrial wall thicknesses were very similar, with values measured on RMI between 0.9 and 1.5 cm for the LV and between 1 and 3 mm for the LA. The limited spatial resolution of the Gemini PET makes it unfit to study the atria metabolism as the atria walls and interatrial septum are hardly seen on the reconstructed images (see figure 4). On the other hand, these structures are already visible on the Vereos reconstruction without RR, and become sharper when using a RR algorithm. The manufacturer reconstruction including the PSF RR still provide a better image quality compared to post-reconstruction RR (see figure 4 and table 2).
Unlike the cardiac phantom filled in the present study with a homogeneous activity distribution in the walls, patients may suffer from pathologies inducing heterogeneities in RA or LA uptake. However, resolution recovery methods tend to reconcentrate the detected activity in the regions where it originates. If heterogeneities larger than the spatial resolution of the PET system are located in the RA or LA walls, they will still be visible after RR.
One limitation of the present study is that patient motions, like breathing and cardiac motions, were not taken into account. The heart motion impact on the present results is difficult to evaluate. Using cardiac gating acquisition would help the quantification with the RR method applied on each gated bin separately, and afterwards summing together the activities of the bins in order to recover the non-gated statistics. Without gating the RA and LA walls on the PET image are enlarged by the heart motion. The RR cannot correct for that motion but it will still bring back the activity in that enlarged wall, improving the quantification as anyway in that case the VOIs will have to be enlarged too to take the motion blurring into account.
As the reconstruction algorithms with RR, seen as black-boxes by the end user, might strongly differ between manufacturers, the optimized parameters found in this study are only adapted to the Philips Vereos TOF-PET/CT. The observed trends for the activity recovery and for the SNR convergences would probably be similar for other systems but this needs to be checked per TOF-PET/CT system in order to extract the best set of reconstruction parameters.