This study addressed the relationship between contrast recovery and the COV under conditions of low counts while applying a BPL algorithm. The investigated BPL algorithm Q.Clear by GE Healthcare is available on their newer Discovery PET/CT scanners. BPL algorithms have been investigated before by other groups, however, not combining measurements of a NEMA and small sphere phantom while varying imaging acquisition times per BP. In addition, studying the correlation between COV and detector sensitivity within overlapping bed positions is new. In clinical PET/CT imaging, the choices to be made when optimizing scanning protocols depend on the primary goal: whether it is to achieve shorter imaging times, decreased radiation dose, or improved image quality and lesion detectability. In practice, it will always be a trade-off between these factors. Our study confirmed that in order to have similar or better noise reduction, the β-value of the BPL algorithm needs to be higher than 450 to have an optimal result, in particular under conditions of low counts. In addition, our results suggest to use a BPL algorithm with an optimized β-value, instead of a traditional OSEM+PSF algorithm.
To support this, we compared the application of a BPL algorithm to OSEM with and without PSF in phantom as well as in patient studies. As reference to start our optimization the scanning and post-processing settings recommended by the manufacturer were used. Our study confirmed that PSF for OSEM indeed has a positive effect on SUV recovery and COV. For BPL, where the only user adaptable setting is the β-value, we showed that increasing the β from 450 to 550 and 700 reduces noise. However, this comes with somewhat less small lesion detectability and lower SUV recovery values (Figs 1-2). Other studies showed that lowering β-values down to 100 has a positive effect on contrast recovery though with higher noise levels, in general they conclude that for 18F-FDG whole body scans a β-value of 400-500 would be optimal [14, 16, 27]. We showed that SUV recoveries achieved using BPL with β-values of 450 to 700 always exceed those achieved by OSEM with or without PSF. In support of our findings others already showed that SUV recoveries achieved when optimizing OSEM+PSF do never reach the same levels as achieved using BPL . With phantom studies with spheres sized 4-37 mm, we demonstrated improved small lesion detectability for spherical lesions down to 5 mm in diameter due to high levels of convergence achieved with BPL, which extends the findings of an earlier phantom study with spheres sized 10-37 mm . In addition, we show that this effect only holds for medium (4-to-1) and high (10-to-1) sphere-to-background ratios. Moreover, this contrast enhancement effect was also seen for patient lesions, which complies to earlier studies [14,15,16]. The improved detectability by BPL of lesions sized 8-13 mm seems of special clinical interest, because these lesions would be sometimes missed using OSEM+PSF. In addition, spill-in or spill-out of counts is also effectively reduced using BPL, as demonstrated by lower SUV levels seen in the lung insert in the phantom, which is in line with an earlier study by Teoh et al. . Therefore, we expect in the coming years that, besides hardware developments like digital TOF, post-processing developments like the implementation of BPL algorithms will enhance the performance of PET at a similar pace as we have seen the last decade. Likely, the EANM will have to continue updating the EARL min-max accreditation limits to keep up with these developments, as they did over the last decade [8, 9].
Within Europe, as well as in many countries outside Europe, radiation dose to the patient is an important optimization factor. Therefore the activity administered to the patient is limited, which next to being beneficial for dose reduction, is also economically beneficial due to lower costs for radiopharmaceuticals. On the other hand, the total imaging time per patient study should be limited for patient comfort and optimal scanner throughput. Optimizing acquisition time per bed position for different anatomical regions is common practice for whole body PET/CT studies. For example, when scanning the legs, multiple bed positions are needed. However, attenuation is much less than in the torso, and PET images of the legs have, in general, less clinical value. Therefore, acquisition times per bed position for the legs section are, at many institutes, set a factor 1.5-3× lower than for the rest of the body, which brings down the total scan time per patient with 5-10 minutes. On the other hand, images should still be of a quality high enough to see anomalies with enough detail, and noise levels should be acceptable and show no artefacts.
SNRs in PET images depend on the number of desintegrations detected by the crystals in individual voxels of the PET/CT detector. These detectors have a 3D sensitivity profile that linearly drops from the center, thus COV at the edges of the axial field-of-view (FOV) will be higher than in the center. To improve counting statistics, and thus image quality, BPs need to overlap. The Poisson noise at a voxel level relates to the detected counts (N) as 1/√N. For BP overlap regions noise would correlate with accumulated counts, and thus, the COV per slice correlates to the sensitivity profile of overlapping BPs. The Discovery 710 scanner has an default overlap of 23%, as recommended by the manufacturer. In our studies at 1 min/bed, we observed the appearance of bands in the PET images. These bands have a 1.3× higher COV at the edges of the overlap region than in the center of the overlap region (average local sensitivity 83%), and correlates to regions where the local sensitivity is 50%. The default choice of a 23% overlap, as recommended by the manufacturer, was made due to the axial FOV of the system of only 15.7 cm; several other systems have a longer axial FOV . For patient comfort, it was decided to have an overlap smaller than 50% and thus less BPs, instead of more overlap and more BPs. Most often, other vendors advise to use a higher overlap up to 50%. Under these conditions, overall scan time per patient can be maintained by reducing acquisition time per BP . Another benefit will be reduced differences between center and edge positions and thus result in smoother images. Therefore, we suggest to increase the bed overlap or increase the acquisition time per bed in our center.
An alternative to bring down the COV is to increase the β-value; we show that a β of 700 can effectively reduce the COV in the image to acceptable levels. Although at the same time it slightly compromises lesion detectability and SUV recovery. The EARL guidelines state that an optimal COV for exact quantification of the SUV in phantom studies should be below 15% . From Figure 4f we may conclude that to comply with a COV < 15% the scan time per bed (counting statistics) or the administered activity should be increased. When making a whole body scan with variable scan times per bed, β-value modulation to balance SUV recovery and COV would be an option worth considering in future reconstruction software for PET/CT scanners.