Impact of Bayesian penalized likelihood reconstruction on quantitative and qualitative aspects for pulmonary nodule detection in digital 2-[18F]FDG-PET/CT

To evaluate the impact of block sequential regularized expectation maximization (BSREM) reconstruction on quantitative and qualitative aspects of 2-[18F]FDG-avid pulmonary nodules compared to conventional ordered subset expectation maximization (OSEM) reconstruction method. Ninety-one patients with 144 2-[18F]FDG-avid pulmonary nodules (all ≤ 20 mm) undergoing PET/CT for oncological (re-)staging were retrospectively included. Quantitative parameters in BSREM and OSEM (including point spread function modelling) were measured, including maximum standardized uptake value (SUVmax). Nodule conspicuity in BSREM and OSEM images was evaluated by two readers. Wilcoxon matched pairs signed-rank test was used to compare quantitative and qualitative parameters in BSREM and OSEM. Pulmonary nodule SUVmax was significantly higher in BSREM images compared to OSEM images [BSREM 5.4 (1.2–20.7), OSEM 3.6 (0.7–17.4); p = 0.0001]. In a size-based analysis, the relative increase in SUVmax was more pronounced in smaller nodules (≤ 7 mm) as compared to larger nodules (8–10 mm, or > 10 mm). Lesion conspicuity was higher in BSREM than in OSEM (p < 0.0001). BSREM reconstruction results in a significant increase in SUVmax and a significantly improved conspicuity of small 2-[18F]FDG-avid pulmonary nodules compared to OSEM reconstruction. Digital 2-[18F]FDG-PET/CT reading may be enhanced with BSREM as small lesion conspicuity is improved.


Impact of BSREM on quantitative values.
The results of the quantitative analysis including SUV max , SBR, SNR, CBR, and CNR from BSREM and OSEM datasets are given in Table 2. Pulmonary nodule SUV max was significantly higher in BSREM images as compared to OSEM images (p = 0.0001). The same was observed for SBR, SNR, CBR, and CNR, see Table 2. For CBR and CNR, negative values were observed since nodule SUV mean was smaller than the relatively constant SUV mean in the descending aorta. With lower SUV mean in OSEM compared to BSREM reconstruction (2.2 ± 1.85 vs. 3.5 ± 2.9, respectively), such negative values were observed more frequently in OSEM images.
In a size-based analysis, the relative increase in SUV max and other quantitative parameters was more pronounced in smaller nodules (≤ 7 mm) as compared to larger ones (8-10 mm, or > 10 mm, Table 3 and Fig. 1). A graphical illustration of the quantitative impact of BSREM on SUV max with pulmonary nodules stratified by size and activity is given in Fig. 2.
Qualitative values: impact of BSREM on nodule conspicuity. The mean conspicuity score was 2.8 for reader 1 and 2.8 for reader 2 in BSREM, which was significantly higher compared to OSEM (2.3 for reader 1 and 2.2 for reader 2; p both < 0.0001). Figure 3 illustrates the conspicuity score ratings for all nodules (n = 144) of the two readers for BSREM and OSEM reconstruction. Inter-reader agreement for nodule conspicuity with OSEM was substantial (OSEM: Cohen's kappa = 0.747) and for BSREM almost perfect (BSREM: Cohen's kappa = 0.846).
Representative images of a study subject undergoing 2-[ 18 F]FDG-PET/CT for oncologic staging are given in Fig. 4.

Clinical follow up for nodule etiology.
To gain more information about the etiology of the 2-[ 18 F]FDGavid pulmonary nodules, the clinical information system of our hospital was screened for information about the etiology of the nodules, Fig. 5. Overall, 20.1% (29/144) of nodules were found to be malignant, as proven by pathology. Another 84 of 144 nodules (58.3%) were clinically suspected to be malignant (e.g., owing to growth on follow-up imaging), albeit without pathological proof. Only 2.8% (4/144) of nodules were proven by pathol-  www.nature.com/scientificreports/ ogy to be benign. Another 11.8% (17/144) of nodules was assumedly benign, based on radiological follow-up exams. 6.9% (10/144) of nodules remained undetermined, since no or inconclusive follow-up data was available.

Diagnostic performance based on SUV max (BSREM vs. OSEM). Receiver operating characteristic
(ROC) curves evaluating the value of BSREM and OSEM to differentiate malignant from benign nodules based on SUV max are presented in Fig. 6. The area under the curve (AUC) values were 0.639 (p = 0.044) and 0.675 (p = 0.011), respectively, with no statistically significant difference between the two algorithms (p = 0.128).

Discussion
This study sought to evaluate the impact of BSREM reconstruction on the quantitative and qualitative aspects of 2-[ 18 F]FDG-avid pulmonary nodules compared to conventional OSEM reconstruction on a latest-generation silicon-based digital detector PET/CT scanner. The major findings of our study are as follows: (1) BSREM reconstruction algorithm leads to a significant increase in SUV max and other quantitative parameters in small pulmonary nodules compared to OSEM, with an average increase of nodule SUV max by 53%. (2) The quantitative impact of BSREM was most pronounced in the subgroup of smallest nodules (≤ 7 mm), with a mean relative increase in SUV max by 80% in this subgroup. (3) BSREM yielded a higher conspicuity of pulmonary nodules than OSEM. (4) The use of BSREM did not improve the overall accuracy of 2-[ 18 F]FDG PET/CT for differentiating malignant from benign nodules.
Pulmonary nodules are a frequent but unspecific CT finding in the daily routine of radiologists and nuclear medicine physicians 1 . In the Pan-Canadian Lung Cancer Screening Study (PanCan), the reported high percentage (74%) of patients with at least one pulmonary nodule was in contrast to the low percentage (5.5%) of actually malignant nodules 15 . In non-oncological subjects, the guidelines by the Fleischner Society explain how to deal Table 3. Relative changes of maximum standardized uptake value (SUV max ), nodule signal-to-background ratio (SBR), nodule signal-to-noise ratio (SNR), contrast-to-background ratio (CBR), and contrast-to-noise ratio (CNR) in block sequential regularized expectation maximization (BSREM) compared to ordered subset expectation maximization (OSEM) reconstructions as reference. Values are given as mean ± standard deviation.

BSREM vs. OSEM
All nodules, n = 144 Nodules 1-7 mm, n = 46 Nodules 8-10 mm, n = 43 Nodules > 10 mm, n = 55  Figure 1. Size-based analysis of the relative increase in SUV max showed a more pronounced quantitative impact in smaller nodules (≤ 7 mm) compared to larger nodules sized 8-10 mm and > 10 mm, (*p value = 0.05, ****p value = 0.0001). The whiskers of the box plot range from minimum to maximum.  7 . PET/CT was also found useful for personalizing patient management by identifying the "reference" nodule deserving histological examination 7 . For many years, 2-[ 18 F]FDG PET/CT was regarded unsuitable for the assessment of small pulmonary nodules, mainly owing to the comparably low spatial resolution of PET 18 . Due to recent technological advancements, including novel digital detector systems and improved reconstruction algorithms, pulmonary nodules are now detected more frequently on PET . Indeed, in our cohort 89/144 (62%) of FDG-avid nodules were ≤ 10 mm, and 46/144 (32%) were even < 8 mm. In addition to the novel detector system, BSREM further enhances the SNR, SBR, CNR, CBR and SUV max , particularly of small nodules. In our cohort, the average SUV max increased from 3.6 with OSEM to 5.4 with BSREM, which represents an increase by 53%. In a previous study by Teoh et al. using a photomultiplier tube PET system, it was also reported that BSREM increases the SBR/SNR as compared to OSEM in small pulmonary nodules 11 . Similar to Teoh et al., in our study the increase in SUV max of pulmonary nodules in BSREM did not translate into significant differences of ROC curves using SUV max as a single determinant  www.nature.com/scientificreports/ of malignancy. In another study small (< 10 mm) suspected lymph node metastasis had higher SUV max when reconstructed with BSREM compared to OSEM 12 . As a limiting factor of this finding by Economou et al. 12 it needs to be pointed out that they did not only use two different reconstruction algorithms, but also different PET-scanners. The retrospective pilot study by Howard et al. found-besides increased SUV max as a quantitative measure-also increased visual lesion conspicuity (as a qualitative measure) in 32 analyzed nodules that were previously described as "too small to characterize" 13 . Today, the study by Howard et al. is the only hint that sole "quantitative improvement" would also affect lesion conspicuity. Furthermore, it is clear that augmented quantitative accuracy in PET may not consequently translate into an improvement of clinical reading. Therefore, performance assessments of readers were included in our study to complement the quantitative approach and further validate improved conspicuity of small nodules on BSREM. We could show that PET reading may be enhanced with BSREM, since small lung lesion conspicuity was improved in our study. The improved conspicuity on BSREM may be related to the fact that the increase in SUV max and the other quantitative parameters (SBR, SNR, CBR and CNR) translate into a better lesion recognition by the human eye. For the quantitative data it was previously described that BSREM increased quantitation accuracy compared to OSEM, especially in cold background regions, such as lungs 19,20 . Similarly to Teoh et al., we believe that quantitative increases in SUVmax are due to almost full convergence of BSREM, compared to the only partial convergence of OSEM (in  www.nature.com/scientificreports/ our study two iterations were used) 9,11 . Due to the limited convergence of OSEM, the true SUV max is consistently underestimated. The underestimation in OSEM is particularly pronounced in small lesions 21 . However, this underestimation can normally be mitigated if point spread function modeling is used 22 . Interestingly, we were still able to measure differences, although both OSEM and BSREM used point spread function modeling. Moreover, as described in previous phantom studies, BSREM improved the quantification accuracy especially for smaller (i.e., sub-centimeter) nodules 23 . Our study has some limitations. First, our study group is relatively small and limited to a single center. Second, we included patients with 2-[ 18 F]FDG-avid pulmonary nodules without further proof of the etiology of these nodules at the time of inclusion (i.e., malignant vs. benign). However, we feel that in an oncological cohort, any 2-[ 18 F]FDG-positive nodule is potentially relevant and warrants at least follow-up imaging, considering the generally high pre-test probability of malignancy. The relevance of the 2-[ 18 F]FDG-positive nodules was confirmed by our analysis of nodule etiology, since the majority (78.4%) of nodules was either pathology-proven (20.1%) or clinically highly suspected (58.4%) to be malignant (20.1%). Third, we did not reconstruct images with different β-values of BSREM or with different OSEM settings, which may differently affect quantitative or qualitative features of lesions in different subjects, depending for example on the individual 2-[ 18 F]FDG dosage or BMI. It is expected that further iterations using OSEM and PSF may alter quantitative aspects of pulmonary nodules. However, it is well known that such high-iteration OSEM images are deteriorated by noise and are unusable for clinical PET reading.
In conclusion, BSREM results in a significant increase of SUV max and improved signal-to-noise ratio in small 2-[ 18 F]FDG-avid pulmonary nodules compared to conventional OSEM reconstruction. The conspicuity of small pulmonary lesions on digital detector 2-[ 18 F]FDG PET/CT may be enhanced using BSREM reconstruction.  www.nature.com/scientificreports/ of 2-[ 18 F]FDG distribution. The CT scan was acquired using an automated dose modulation technique (range 15-100 mA) with 120 kVp. After the CT scan, PET images were acquired covering the identical anatomical region. The acquisition time for PET was 2.5 min per bed position, with 6-8 bed positions per patient (depending on patient size), with an overlap of 23% (17 slices). The PET was obtained in 3D mode and slice thickness was 2.79 mm. Two PET dataset reconstructions were generated using (1) BSREM (Q.Clear, GE Healthcare) with a default β-value of 450, and (2) OSEM with two iterations, 24 subsets (i.e., 48 image updates, as recommended by the vendor) and 6.4 mm Gaussian filter with time-of-flight reconstruction and point spread function modelling (OSEM; Vue Point FX with SharpIR, GE Healthcare). All PET datasets were reconstructed with a 256 × 256 pixel matrix.

Materials and methods
Quantitative imaging analysis. Quantitative analyses were performed by one reader (*Blinded for Review*, with 2 years of experience in radiology/nuclear medicine NL). A standard volume of interest (VOI) was used to record the maximum standardized uptake value (SUV max ) of each pulmonary nodule in both BSREM and OSEM datasets. Nodule diameter was measured in the long-axis on axial CT slices in lung window. Similar to Teoh et al. 11 , background SUV was recorded in the right lobe of the liver (parenchymal organ background) and within the descending aorta (blood pool background) at the level of the carina, with 4.0 cm-diameter (liver) and 1.0 cm-diameter (aorta) spherical VOIs. Only liver parenchyma appearing normal on both PET and CT was used as a reference. In both backgrounds and for both reconstructions, the mean standardized uptake value (SUV mean ) and the standard deviation of the standardized uptake value (SUV SD ) within the VOIs were recorded. As previously described, a signal-to-background ratio (SBR) based on these measurements was calculated for each nodule, defined as the lesions' SUV max divided by the SUV mean in the descending aorta 10 . The liver SUV SD served as a measure of noise. Nodule signal-to-noise ratio (SNR) was defined as the lesion's SUV max divided by the liver SUV SD . Furthermore, the (nodule SUV mean minus the SUV mean in the descending aorta) divided by the SUV mean in the descending aorta defined a calculated contrast-to-background ratio (CBR) 10 . Lastly, contrast-tonoise ratio (CNR) was calculated, defined as the (nodule SUV mean minus the SUV mean in the descending aorta) divided by the liver SUV SD Clinical follow-up for nodule etiology analysis. The clinical information system was screened for the best available data on the etiology of the 2-[ 18 F]FDG-avid pulmonary nodules. Every available information (e.g., pathology reports, radiological follow-up scans, oncology reports) were used to determine the etiology. Nodule etiology was grouped into five categories as follows: (1) malignancy proven by pathology; (2) malignancy to best clinical knowledge; (3) benign lesion proven by pathology; (4) benign lesion to best clinical knowledge; and (5) no or inconclusive follow up available.
Diagnostic performance of SUV max (BSREM vs. OSEM). The diagnostic performance of BSREM and OSEM to differentiate malignant from benign nodules based on SUV max was assessed using an ROC curve and AUC values.
Statistical analyses. Categorical variables are expressed as proportions, and continuous variables are presented as mean ± standard deviation or median (range), depending on the distribution of values. Wilcoxon matched pairs signed-ranks test was applied for comparison of SUV max , SBR, SNR, CBR, and CNR values in BSREM vs. OSEM. Furthermore, the same test was also used to compare the subjective analysis of the readers (i.e., conspicuity score per nodule). Mann-Whitney U test was performed for size-based comparisons of relative increase of SUV max . For the post-hoc analysis, a Bonferroni-corrected p-value of 0.025 (0.05/number of tested size groups) was considered to indicate statistical significance. The difference ratios of SUV max , SBR, SNR, CBR, and CNR values in BSREM vs. OSEM datasets were calculated using the results of OSEM reconstructions as reference as follows: (variable in BSREM reconstruction minus variable in OSEM reconstruction)*100/(variable in OSEM reconstruction). All analyses were performed with statistics software (SPSS version 26.0, IBM Corporation, Armonk, NY or GraphPad Prism version 8.3.1). A two-tailed p-value of < 0.05 was considered to indicate statistical significance. www.nature.com/scientificreports/ Publisher's note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Data availability
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http:// creat iveco mmons. org/ licen ses/ by/4. 0/.