Phantom and clinical assessment of small pulmonary nodules using Q.Clear reconstruction on a silicon-photomultiplier-based time-of-flight PET/CT system

To evaluate the quantification accuracy of different positron emission tomography-computed tomography (PET/CT) reconstruction algorithms, we measured the recovery coefficient (RC) and contrast recovery (CR) in phantom studies. The results played a guiding role in the partial-volume-effect correction (PVC) for following clinical evaluations. The PET images were reconstructed with four different methods: ordered subsets expectation maximization (OSEM), OSEM with time-of-flight (TOF), OSEM with TOF and point spread function (PSF), and Bayesian penalized likelihood (BPL, known as Q.Clear in the PET/CT of GE Healthcare). In clinical studies, SUVmax and SUVmean (the maximum and mean of the standardized uptake values, SUVs) of 75 small pulmonary nodules (sub-centimeter group: < 10 mm and medium-size group: 10–25 mm) were measured from 26 patients. Results show that Q.Clear produced higher RC and CR values, which can improve quantification accuracy compared with other methods (P < 0.05), except for the RC of 37 mm sphere (P > 0.05). The SUVs of sub-centimeter fludeoxyglucose (FDG)-avid pulmonary nodules with Q.Clear illustrated highly significant differences from those reconstructed with other algorithms (P < 0.001). After performing the PVC, highly significant differences (P < 0.001) still existed in the SUVmean measured by Q.Clear comparing with those measured by the other algorithms. Our results suggest that the Q.Clear reconstruction algorithm improved the quantification accuracy towards the true uptake, which potentially promotes the diagnostic confidence and treatment response evaluations with PET/CT imaging, especially for the sub-centimeter pulmonary nodules. For small lesions, PVC is essential.


Results
Phantom. From the visual assessment ( Fig. 1), as more advanced reconstruction techniques (from OSEM to Q.Clear reconstruction) were introduced into the PET reconstruction, the overall image quality improved. Thanks to the Q.Clear algorithm, the smallest sphere (diameter = 10 mm) was particularly outstanding with clearer boundaries; meanwhile, reduced background noise was also achieved compared with the other reconstruction methods. Table 1 and Fig. 2 has shown that A m (Bq/mL), CR (%), and RC (%) increased gradually in the order from OSEM, TOF, TOF-PSF to Q.Clear. The average A m and RC of 5 spheres (10 mm-28 mm) using the Q.Clear algorithm were significantly higher than those of other reconstruction methods (P < 0.05), while the biggest sphere (37 mm) showed no significant difference among reconstruction methods (P > 0.05). The CR values of all hot spheres using the Q.Clear algorithm had higher values than those of the other reconstruction algorithms (P < 0.05) ( Table 2). The RC values of 6 hot spheres calculated from the Q.Clear algorithm reached up to 59.0% Clinical characteristics. An illustration of the same pulmonary nodule (long-axis diameter 7.2 mm) on the PET images reconstructed using different reconstruction algorithms was shown in Fig. 3. The lesion became clearer when the Q.Clear reconstruction was used. The SUVmax and SUVmean calculated from all the small pulmonary nodules (n = 75) increased gradually in the order from OSEM, TOF, TOF-PSF to Q.Clear and SUVs Figure 1. Comparison of PET reconstruction methods for the NEMA phantom with 6 spheres (diameter 10-37 mm) filled with 13.2 kBq/mL Fluoride ions in a 4-to-1 contrast ratio. The mean uptake values (Bq/mL) of the 6 spheres (in order from 10 to 37 mm) for 4 reconstructions were as follow: 4.12, 5.66, 6.57, 7.52, 9.03, 9. www.nature.com/scientificreports/ determined from the Q.Clear algorithm were significantly higher than those from other reconstruction methods (P < 0.001 for SUVmax and SUVmean) (Fig. 4). However, the significance was highly affected by the size of nodules, as shown in Table 3. SUVmax and SUVmean determined from sub-centimeter pulmonary nodules had highly significant differences among different reconstruction techniques (P < 0.001), while the medium-size group showed no significant difference (P > 0.05).    Fig. 6. The PVC %ΔSUVmeans were 160.1%, 138.4%, 116.4%, and 75.1% for OSEM, TOF, TOF-PSF and Q.Clear, respectively (P < 0.001). A highly significant difference existed not only between the original SUVmean and the corr SUVmean in each reconstruction algorithm (P < 0.001 for all) but also between the Q.Clear algorithm and the other three ones in corr SUVmean (P < 0.05 for all). However, differences in SUVs measured from four   www.nature.com/scientificreports/ reconstruction methods were not significant for the medium-size group (Table 3, P = 0.649 for SUVmax and P = 0.559 for SUVmean).

Discussion
A reliable and precise measurement of radiopharmaceutical uptake is more and more important in PET, particularly for differential diagnosis, treatment planning, and therapy response evaluation. It is still challenging to achieve accurate quantification of radiotracer uptake in the small lesions due to the partial convergence in current reconstruction algorithms and the partial volume effect, which usually underestimate SUV. In this study, we demonstrated the values of using the Q.Clear reconstruction algorithm on the SiPM-based PET/CT platform for improving quantification accuracy and image quality. The Q.Clear reconstruction algorithm revealed considerable enhancement towards true uptake compared with the other reconstruction methods, especially in sub-centimeter pulmonary nodules. In this study, image reconstruction parameters were chosen based on the routine clinical protocols in our hospital. Small lesions are more clinically challenging. It is hard to reach full convergence on the routine OSEM based algorithms and is heavily affected by PVE. Nodules with a size under 25 mm were considered as small lesions. The Q.Clear-reconstructed images showed the highest RC and CR compared to the other three methods (Fig. 2), suggesting the Q.Clear reconstruction algorithm improved the quantification accuracy of PET imaging approaching the true uptake. However, it was found that the RC difference among the four reconstruction methods was not significant (P > 0.05) in the largest sphere (37 mm). The reason for this is, under the same number of iterations, the degree of convergence of OSEM based reconstruction algorithms was higher in large subjects, which can be seen in Fig. 2 that larger spheres always had higher RC values compared with smaller spheres regardless of reconstruction methods used.
For the quantification of pulmonary nodules, significant increases in the SUVmax and SUVmean were observed when reconstructed with the Q.Clear algorithm compared to other three methods (Fig. 4), and the increases in SUVs for those of sub-centimeter nodules were significantly greater than those of the medium-size group ( Table 3, P < 0.001). By comparing the changes of SUVs among different reconstruction methods in the medium-size group, increments can be found but were not statistically significant (Table 3, P > 0.05). This was because, firstly, the medium-size group was closer to full convergence on routine OSEM based algorithms compared to the sub-centimeter group; secondly, the number of nodules enrolled in the medium-size group was inadequate (n = 18) to reach statistical significance. Similarly, Teoh et al. also showed that the greatest improvement was found in malignancy detection in nodules ≤ 10 mm compared with larger ones 21 . This concluded the Q.Clear algorithm exerted more profound impacts on the SUV measurement of small pulmonary nodules, which potentially facilitated the enhancement of the lesion visibility and detectability, especially for sub-centimeter lesions.
The partial volume effect is another impacting factor on SUV quantification accuracy of small lesions, which also underestimates the true uptake. In our phantom study, RC was calculated to reveal the direct relationship between the measured and true radioactivity. Then the association between RC values and sphere diameter measured on CT images was established for each reconstruction method. The linear regression was performed, and good regression coefficients were achieved (Fig. 5). Based on the equations and the measured nodule size in CT images, we corrected the partial volume effect to improve the SUV measurement accuracy of all pulmonary nodules. This RC-based PVE correction approach is on assumptions that the lesion has a regular spherical shape and uniform distribution of radioactivity 25 . Due to the limited size range of phantom spheres being from 10 to 37 mm, the PVE correction for those nodules either smaller than 10 mm or larger than 37 mm should be careful. In addition, in Fig. 6, the number of outlier values decreased in the order from OSEM, TOF, TOF + PSF to Q.Clear, which indicated the improvement in the SUV reliability. A few previous studies have demonstrated this method for PVE correction 11,26,27 25 . Our finding on the relationship between RC and sphere diameter showed consistency with their result. Future elaborated phantom experiment designs should be dedicated to investigating different clinical imaging scenarios with different lesion sizes, lesion shapes, image contrasts, and radioactivity levels. It should be noted that this PVE correction method is equipment-specific, which means different PET scanners may have different RC-diameter behaviors.
The respiratory motion during PET imaging may compromise the spatial resolution, resulting in smearing or blurring effects of PET images, which would eventually affect SUV measurement. Existing researches suggested that respiratory movement leads to overestimations of radiotracer-avid target volume and reduction in the SUV due to the recorded number of coincidence events tending to distribute in a larger volume 28 . A data-driven respiratory gating function is now available on the GE system, which has shown out-performance than the devicebased gating 29 . The respiratory gating system can be employed in a future study to facilitate the improvement of spatial resolution and quantification accuracy, especially when radiotherapy treatment planning is desired [30][31][32] .
Furthermore, the small pulmonary nodules in our study were not histopathologically verified. The relationship between malignancy and SUV measurement of pulmonary nodules under Q.Clear reconstruction need to be investigated, which could benefit differential diagnosis based on quantitative PET/CT imaging.

Conclusion
In our phantom and clinical studies, the Q.Clear reconstruction algorithm combined with SiPM-based digital PET/CT platform significantly improved the quantification accuracy towards the true uptake by accessing RC and CR values, which potentially promotes the diagnostic confidence and treatment response evaluation with PET/CT imaging, especially for the sub-centimeter pulmonary nodules. For small lesions, the PVE correction is essential. where A M is the measured activity (in kBq/mL) in each sphere delineated on CT images; A B is the measured activity in the background. A K is the known activity (in kBq/mL) in the sphere; C is the known ratio of activity in the sphere to the background (that is 4:1 in the study).
PET/CT imaging. 18  where x is the image estimate; i is the pixel index; y i represents the measured PET coincidence data; P is the system geometry matrix; β is the penalization factor; R(x) is the penalty to control noise. After a large number of experiments in our department, 350 is the best beta setting, which is used in our clinical practice.
Clinical evaluation. The   PVC for pulmonary nodules. A logarithmic model was generated based on the relationship of RC and sphere diameter (measured on CT images) in each reconstruction method. The partial volume effect correction was performed on all pulmonary nodules, based on the small-size nature, by dividing the corresponding RC. The PVC percentage increase, %ΔSUVmeans, was calculated as the difference between PVE corrected SUVmeans and original ones (4).
Statistical analysis. Statistical analysis was conducted using SPSS Statistics 25.0 (IBM Co., New York, USA). Kruskal Wallis H test was utilized to analyze differences in phantom data (RC and CR) and in clinical data (SUVs) among the four reconstruction algorithms. Mann-Whitney U test was used to compare the SUVmean difference before and after PVC. P value < 0.05 was considered as statistically significant differences, while P < 0.001 was taken as highly significant differences.