The aim of this study was to investigate the impact of Q.Clear reconstruction methods on brain kinetic modelling analysis by evaluating the performance of Q.Clear, against the performance of OSEM in the presence of a small number of counts, in brain images acquired in a PET-MR system. To our knowledge, we are the first to investigate Q.Clear reconstruction performance for brain kinetic modelling analysis rather than simplified quantification methods like standard uptake value (SUV). We also report here that, for low count brain scans in comparison to whole-body PET imaging, much lower β levels (between 100 and 200) are required to achieve the same quantitative results to those obtained with a OSEM method.
The results for all structures, apart from the Substantia Nigra, appeared to be unaffected by the reconstruction method as the changes in the CV were minimal. The Substantia Nigra however appeared to be vulnerable to the reconstruction method under a low count scenario as not only the results appeared more dispersed, but it was also observed an increase of almost 12% for the CV calculation. This finding demonstrates that, when conducting kinetic modelling with an SRTM, the reconstruction algorithm used may have a different impact on different brain structures. This project did not consider partial volume effect correction which is important for small structures such as the Substantia Nigra. Even though Q.Clear improves spatial resolution versus OSEM due to PSF corrections, this is still limited and a consequence of it is the partial volume effect which can affect the PET images quantitatively [9]. Therefore, the results for the Substantia Nigra could be underestimated by a spill-over effect from the white matter located in the midbrain [23].
The penalisation factor in Q.Clear performs as a noise suppression term with higher β values resulting in stronger noise reduction, whilst preserving edges [24]. This explains the decrease in the mean BPND results with the increase in β value, for the SN, St, and GP. The exception to this can be observed in the thalamus as there is a slight increase in the mean BPND results with the increase in β value, possibly due to the low target density in that region (with BPND values approximately 10 times lower than high density and large regions, such as the striatum).
The BPND obtained from the in vivo data demonstrates that, in a low count scenario, Q.Clear with β100 has the lowest bias when compared to the standard low count OSEM reconstruction for the SN, GP and Th. For the same metric in the Striatum, Q.Clear with β200 has the lowest bias. Furthermore, when the BPND for the Cd and Pt are investigated individually it is also noted that Q.Clear with β200 presents the lowest bias for both structures. These results are further substantiated by the multi comparison results which demonstrate that Q.Clear with β100, 200 and 400 are the only reconstructions across all structures that do not present statistically significant (0.01 < p ≤ 0.05), very statistically significant (0.001 < p ≤ 0.01) or extremely statistically significant (0.0001 < p ≤ 0.001) differences when compared to the standard reconstruction.
Lindström et al. (2017) investigated clinical whole-body scans which were acquired in a GE PET-CT system and reconstructed with Q.Clear and TOF-OSEM. They found that in order to obtain a noise equivalence to TOF-OSEM reconstructions with 3iterations, 16subsets and 5 mm Gaussian filter, a Q.Clear reconstruction with β600 should be performed for radiotracers such as 68 Ga-DOTATOC, 18F-FDG and 18F-Fluoride and a Q.Clear reconstruction with β400 should be performed for 11C-acetate [25].
Scott et al. (2019) aimed at optimising Q.Clear for 90Y quantitative imaging by preparing a National Electrical Manufacturers Association (NEMA) image quality phantom with an 90Y solution and scanning it on a GE Discovery 710 PET-CT scanner. Images were re-binned in the first instance into 15 min frames and, at a later stage, into 30 and 60 min frames and reconstructed with Q.Clear with β values of 1, 400, 800, 1000, 1200, 1400, 1600, 1800, 2000, 3000, 4000 and 8000. They calculated activity recovery and found that the optimal value for quantification was β 1000, as the reduction in image noise provided by this reduction does not affect quantification [26].
These reports demonstrate that the optimal β value is dependent on the tracer and the OSEM parameters used for a given application (e.g. brain PET studies versus whole-body PET). Notably, brain PET imaging requires more resolved images and this can be achieved with either an OSEM reconstruction with a high number of iterations and subsets or a Q.Clear reconstruction with low β values. It is encouraging that our results are in line with the report by Ross [27] who reconstructed two 18F-FDG brain image datasets with OSEM 3iterations, 32 subsets and 2.5 mm filter and Q.Clear with β150 and found that this β level produced excellent contrast and image quality in both datasets. Reynés-Llompart et al. (2018) evaluated phantom and brain and whole-body patient images which had been acquired in a GE Discovery IQ PET-CT system and reconstructed with Q.Clear with β from 50 to 500. They used various acquisition times to mimic different counts—the 15 s acquisition in their study yielded 19 ± 4 million counts, which represents the closest statistics to the ones mimicked in our study. At a 15 s acquisition and using a lesion to background ratio of 2:1, a β value of 150 maximises the contrast to noise ratio (CNR) for a sphere of 10 mm, a β value of 200 maximises the CNR for a sphere of 13 mm and a β value of 250 maximises the CNR for a sphere of 17 mm. Although in kinetic modelling spatial resolution is of more importance than CNR, it is important to note that β values of this range yield good contrast for small structures. Their results also demonstrated that for images of the torso, the optimal β value would be between 300 and 400, whereas for the brain images, it would be between 100 and 200, which is in line with our observation [28]. This suggests that, unlike diagnostic whole-body studies, using 18F-MISO and/or 18F-FAZA in hypoxic lung lesions [13] and 18F-FDG PET-CT in pulmonary nodules [10], where the optimal β value is 350 and 400 or studies using 68 Ga-PSMA and 18F-Fluciclovine in pelvic lesions [29, 30] which found that the optimal β value was between 400 and 550 and 300, respectively, for brain studies the optimal β value is lower, particularly when accurate quantification is paramount. In fact, phantom and clinical studies conducted with the aim of improving spatial resolution rather than for diagnostic purposes have reported that Q.Clear with low beta values provides better spatial resolution in small structures. Rogasch et al. (2020) investigated image metrics such as spatial resolution, contrast recovery and SNR in phantom images reconstructed with Q.Clear and OSEM with PSF modeling. The team found that when using Q.Clear with β 150 and a high signal to background ratio, the spatial resolution obtained is superior to that obtained when reconstructing images with PSF modelling and/or time of flight [9]. Similarly, a publication by Howard et al. [14] investigating Q.Clear in small pulmonary nodules reported that Q.Clear with a β value of 150 improved visual conspicuity of nodules of approximately 1 cm.
Our work follows a similar approach to that of Teoh et al., Ter Voert et al. and Teoh et al. [10, 29, 30]. However, whereas these investigations were performed in whole-body imaging and focusing on the effect of the algorithm on SUV metrics, our work was performed in quantitative dynamic brain imaging and demonstrates the effects on BPND. To our knowledge, this has not been attempted before. Moreover, our work further sustains the initial observations presented by Reynés-Llompart et al. [28].
A limitation related with the use of Q.Clear that was noted in this study was that for frames with low counts, spurious high counts were seen in the reconstruction and three of the initial frames had to be removed (as was described in the Results section). This demonstrates the importance of the quality control stage in image analysis.