Our study demonstrated that the BPL reconstruction algorithm significantly increased the SUVmax, SUVmean, and SBR while significantly decreasing the MTV of malignant tumor lesions compared with the OSEM and OSEM + TOF reconstruction algorithms. The changes were more obvious in small tumor lesions or lesions with relatively hypometabolic tumors.
[18F]FDG PET/CT is widely used in clinical practice for the quantitative analysis of malignant lesions such as diagnosis, treatment monitoring, and prognosis assessment. Signal acquisition and reconstruction algorithms have the greatest impact on the accuracy of quantitative parameters. Currently, the OSEM algorithm using repeated iterations remains the primary method used for the reconstruction of PET/CT imaging. With increasing iterations, focal maxima of activity further increases, and may overestimate the true activity, as well as image noise increases. To limit the image noise, OSEM data are usually smoothed with a filter. However, this decreases contrast recovery of “hot” lesions which may result in underestimation of the true activity. Furthermore, OSEM algorithm may not be fully convergent. Based on this context, combining or replacing OSEM with other reconstruction algorithms may improve the accuracy of quantitative parameters of tumor lesions’ [18F]FDG uptake. BPL is a reconstruction method that can achieve more appropriate convergence through more iterations, and control the image noise at a suitable level [3]. The noise suppression function can be controlled by adjusting the penalization factor (β), which is the only variable that allows user to input. Reducing the penalization factor (β) increases the contrast recovery coefficients and the background variability. In a NEMA phantom research by Teoh et al., a higher contrast recovery coefficient was achieved by BPL reconstruction than those by OSEM + TOF and OSEM + PSF; that is, the accuracy of the quantitative parameters was improved by the BPL algorithm [1, 10, 11]. And, the signal-to-noise ratio can be improved by BPL at equal contrast recovery [12]. These advantages of BPL reconstruction can increase the accuracy of [18F]FDG quantitative parameters without compromising the image quality compared with the OSEM method, with a particular improvement in small tumor lesions.
Our study compared the [18F]FDG standard uptake parameters and volume metabolism parameters of malignant tumor lesions in four PET/CT reconstruction algorithms. The results showed that the BPL reconstruction algorithm, which based on TOF and PSF, significantly increased the standard uptake parameters (SUVmax, SUVmean, and SBR) of [18F]FDG in tumors, while MTV was significantly decreased. And, these standard uptake parameters were significantly higher in BPL than in OSEM, OSEM + TOF, and OSEM + TOF + PSF by pair-wise comparisons. But, MTV in BPL was significantly lower than OSEM and OSEM + TOF. Murphy et al. compared the effects of BPL and OSEM reconstruction algorithms on [18F]FDG metabolic parameters in malignant solitary pulmonary nodules in the context of non-small cell lung cancer, mediastinal metastatic lymph nodes, and colorectal cancer with liver metastases [6, 7, 13]. Their results showed that the SUVmax of corresponding lesions reconstructed by the BPL were 8.3 (3.2–13.4), 7.0 (1.3–25.3), and 11.6 (2.6–25.7), respectively, which were significantly higher than those by OSEM (P < 0.001). Similarly, Reynés-Llompart et al. compared the [18F]FDG metabolic parameters SUVmax and SUVmean of tumor lesions between PSF and BPL reconstruction algorithms. Their results were consistent with our findings, but no significant difference in mean lesion volume was found between these two algorithms. This may be for the reason that the convergence function was affected by different size lesions when using BPL algorithm [11]; the research objects were lung nodules in the study by Reynés-Llompart’s et al., whereas the range of lesion size was wide in our study. Matti et al. noted that the BPL reconstruction algorithm can improve the accuracy of SUVmax and SUVmean of [18F]FDG in hypermetabolized lesions without changing the metabolic parameters of background tissue, thus resulting in the improvement of SBR of tumor lesions and image quality [12]. However, some literature has suggested that BPL only increases the apparent [18F]FDG metabolic parameters and may not improve the ability to distinguish benign and malignant lesions [6, 7]. When lesions had low SUV, [18F]FDG PET/CT often could not accurately evaluate true metabolic activity, and when both non-malignant and malignant lesions had a high degree of [18F]FDG uptake, [18F]FDG PET/CT was not specific for distinguishing the nature of nodular lesions. Thus, the benefit of using BPL is mainly to provide better lesion visibility and more accurate quantitative parameters for clinical practices.
In this study, we found that the different rates regarding %ΔSUVmax, %ΔSUVmean, %ΔSBR, and the absolute values of %ΔMTV between the BPL and OSEM + TOF were significantly higher than those between the BPL and OSEM + TOF + PSF. Similar results were observed between OSEM and OSEM + TOF + PSF as compared to BPL. Furthermore, %ΔSUVmax, %ΔSUVmean, and %ΔSBR in the OSEM, OSEM + TOF, and OSEM + TOF + PSF groups were significantly negatively correlated with lesion size and the degree of lesion FDG uptake, suggesting that the BPL reconstruction technology had a more significant convergent effect on small lesions and low degree [18F]FDG uptake lesions. Consequently, by increasing metabolic parameters and the detection sensitivity of small lesions, BPL may help find out the small lesions with low degree [18F]FDG uptake and provide further guidance for tumor staging, evaluation, and treatment decisions in clinical practice.
Consistent with a study by Parvizi [5], our results also confirmed that BPL reduced the MTV of tumor lesions. There are two possible reasons, one of which is the effective convergence by BPL that PET images with lower noise and higher contrast through image noise controlling, edge-preservation, and edge artifacts suppressing [14] can increase the SUVmax of a measured lesion, the other reason may be the different methods used for segmentation, as the automatic outline of MTV was based on the lesion’s SUVmax. BPL increased the SUVmean while decreased the MTV of tumor lesions, which results in no significant change in its TLG. Yamaguchi et al. confirmed that the BPL reconstruction algorithm is a feasible method for the suppression of edge artifacts deriving from PSF correction [14], because the edge suppression included in the BPL reconstruction allows suppression of excessive noise that would otherwise develop with high numbers of iterations. By limiting this noise, more iterations can be performed to enable full convergence of the algorithm. In our study, the maximum number of iterations in the BPL reconstruction as opposed to OSEM was 25, which can achieve full convergence of the algorithm. However, the suppression effect of BPL on noise is affected by sphere-to-background ratios and sphere size. When reconstructed by BPL, the suppression of edges is the most obvious and the boundary is sharpest when images of 10 mm spheres are at a high SBR and without background [14]. Additionally, our study further clarified that %ΔMTV was significantly positively correlated with the lesion size and the degree of the lesion’s uptake. However, due to the partial volume effect caused by the limited spatial resolution of conventional PET systems, radiotracer uptake is usually underestimated when lesions are three times smaller than the spatial resolution [15]. The partial-volume effect correction with a recovery coefficient method is important for improving the quantification accuracy of further clinical evaluations.
[18F]FDG-PET/CT quantitative parameters (such as SUVmax, SUVmean and MTV) as imaging metabolic indicators have unique advantages in clinical diagnosis, staging, post-treatment staging, and prognosis assessment of patients with malignant tumors [16,17,18]. The BPL reconstruction algorithm can improve the accuracy of quantitative parameters through effective convergence and promotes its utility as a biomarker of tumor metabolism. Meanwhile, accurate image segmentation is necessary for proper disease detection, diagnosis, treatment planning, and follow-up. Many PET-based automatic segmentation methods such as thresholding-based, stochastic and learning-based, region-based, boundary-based, and multi-modality methods have been proposed [19], but no consensus has been reached on the optimal delineation method. In this study, we chose the thresholding-based PET image segmentation method, which converts a gray-level image into a binary image by defining voxels greater than some value as the foreground and other voxels as the background [19]. The most frequent thresholding value used in the clinical setting is 40–43% of SUVmax, and we adopted a 42% thresholding value in the current study. Recently, a few more advanced PET image segmentation methods were studied, which are less sensitive to SUVmax variations such as: (1) a method that uses fully convolutional networks with auxiliary paths and uses dual-modality PET-CT images to achieve automatic segmentation of nasopharyngeal carcinoma on PET-CT images [20], which can achieve better performance than existing methods based only on CT images and purely fully convolutional networks; (2) a fully automatic and operator independent method based on an extension of the random walk algorithm for the BTV delineation of brain metastases for Gamma Knife treatments, which has the advantage of automatically identifying target and background random walk seeds and has an adaptive threshold to discriminate target from background voxels [21]; and (3) a new fully three-dimensional methodology for tumor delineation in functional images based on active contours and a slice marching approach, which are an active surface defined in the three-dimensional space and can segment all PET slices at once. The algorithm reduces the need for manual input to a minimum and produces tumor segmentations that are independent from the initial input, thus making the result extremely robust and repeatable [22]. However, these image segmentation methods are implemented on the basis of software application. As a clinician, it is difficult to make such a software application, and its clinical application is complex. Although the thresholding-based PET image segmentation method is vulnerable to SUVmax variations, it is easier to implement in clinical practice. Furthermore, based on this method, we obtained meaningful clinical results.
There are several limitations in the current analysis. Firstly, the number of lesions from different tumor entities is relatively small that the results obtained need to be verified in a larger cohort. Secondly, only data from clinical scans are presented and the ground truth for SUV and MTV is lacking, which leads to the uncertainty of diagnostic efficacy of these parameters. Thirdly, we only examined one set of reconstruction parameters for each of the algorithms although previous studies have shown that the differences between algorithms are dependent both on the settings for the OSEM-based algorithm as well as the beta parameter used with the BPL reconstruction [23]. Last but not the least, the thresholding-based PET image segmentation method used in our study is affected by SUVmax variations and may need considerable adjustments for different PET images.
