1. Introduction
Lung cancer continues to be the tumor disease with the leading number of cancer deaths worldwide [
1]. Precise staging is essential for the initiation of adequate therapy [
2]. PET/CT with the glucose analog [
18F]Fluorodeoxyglucose ([
18F]FDG) assumes a central function for staging lung cancer, according to international guidelines [
3,
4]. [
18F]FDG-PET is generally performed as a static scan, at a defined uptake time of 60 to 90 min after intravenous (i.v.) tracer application. However, due to increased [
18F]FDG affinity in inflammatory tissue, [
18F]FDG-PET is known to have limited specificity for an accurate evaluation of thoracic lymph nodes, especially in the presence of frequently associated tumor inflammatory pulmonary disease. Thus, [
18F]FDG-avid lymph nodes must be biopsied before surgery or radiotherapy to rule out malignancy histologically [
3,
4]. However, such an intervention is often difficult and risky in clinical practice due to the often-limited cardiopulmonary reserve. Furthermore, the evaluation of indeterminate lung lesions, which cannot be biopsied due to their location or unfavorable risk–benefit to the patient, is also an indication for PET [
3,
4].
One way to generate complementary PET information is to quantify the tracer distribution over time. Until recently, this was typically feasible using two workflows with significant limitations. The first option is a dynamic acquisition, where the tracer distribution is continuously measured in a defined but limited anatomical region. Using this method, the axial field of view of current well-established PET scanners (generally between 15 and 30 cm) limits the anatomical coverage, which in turn restricts the dynamic acquisition of whole-body data [
5,
6]. A second option is a dual-/multi-time-point PET: this technique combines two or more static PET examinations and calculates the difference in [
18F]FDG uptake [
7,
8,
9]. Whereas traditional dynamic PET is not suitable for whole-body staging due to the limited FOV of the PET scanner, dual-time-point imaging has already shown significantly increased accuracy for the assessment of mediastinal lymph node metastases (LNM) in a large meta-analysis of 654 patients with non-small cell lung cancer (NSCLC) [
7].
Dynamic whole-body PET data can be produced using an innovative combination of dynamic acquisition at the start of the scan followed by multiple subsequent whole-body scans either in the “step-and-shoot” or in the “continuous-bed-motion” technique. This form of dynamic data acquisition can be used for Patlak kinetic modeling, which enables the assessment of [
18F]FDG distribution in different compartments separately for each organ and tissue in the body [
10,
11,
12,
13]. However, the clinical benefit of this technique and of dynamic information on tumor staging has not been completely elucidated.
Therefore, the aim of this prospective study was to assess the feasibility of dynamic whole-body PET acquisition in a clinical setting and to evaluate the diagnostic performance of parametric imaging in the classification of indeterminate lung lesions and lymph nodes.
2. Materials and Methods
2.1. Study Design
Thirty-three consecutive patients with indeterminate pulmonary lesions and a clinical indication for [
18F]FDG-PET/CT were enrolled into this prospective unicentric trial between June 2019 and April 2022, as shown in detail in the Consolidated Standards of Reporting Trials (CONSORT) flow diagram (
Figure 1). This prospective trial was approved by the Institutional Review Board (registry No. 333/2019BO2) and is listed in the German Clinical Trial Register (DRKS-ID: DRKS00017717). All patients signed an informed consent.
2.2. PET/CT Examination Protocol
Patients were asked to fast for at least 6 h prior to examination. Weight, size, and blood sugar level were measured before i.v. tracer administration. Blood glucose level was below 140 mg/dL in all patients without the administration of insulin 8 h prior to tracer application. [18F]FDG dosing was weight-based using 4.0 ± 0.6 MBq/kg. All patients were positioned with arms up on a vacuum mattress on the PET/CT (Biograph mCT, Siemens Healthineers) table to reduce motion artifacts and were asked to breathe as calmly and steadily as possible.
Before PET, a full diagnostic CT with adaptable tube voltage and tube current (CARE KV 120–140 kV, CARE Dose 4D 40–280 mAs) was performed. An iodinated contrast agent (80–100 mL Ultravist® 370, Bayer Vital GmbH, Leverkusen, Germany) was administered to all patients except for contraindications.
The dynamic PET acquisition started simultaneously with the i.v. injection of [
18F]FDG and lasted a total of 80 min. The initial table position was centered over the cardiac region (BI ≈ 6 min) to acquire the individual input function followed by whole-body (WB) dynamic PET of skull to mid-thigh (WB ≈ 74 min) using continuous-bed-motion as described in detail by Karakatsanis et al. and Rahmim et al. [
10,
11,
12].
Image data were subdivided into 43 time frames (12 × 5 s, 6 × 10 s, 8 × 30 s, 7 × 180 s, and 10 × 300 s.) The time activity curve (TAC) was derived by an automatically generated cylindric volume of interest (VOI: 10 mm diameter and 20 mm long) centered in the descending aorta with acquired CT images using ALPHA (automated learning and parsing of human anatomy) as implemented in the vendor’s software (VG70A, Siemens Healthcare GmbH, Erlangen, Germany).
2.3. Reconstruction and Postprocessing
Dynamic PET data (cardiac region and WB) were reconstructed with OSEM 3D reconstruction applying point-spread-function (PSF) and time-of-flight (TOF)—using two iterations, 21 subsets, a 200 × 200 matrix, and a 5 mm Gaussian filter. The reconstructed passes 12–17 of the WB and the resulting TAC were used to perform the Patlak reconstructions with two iterations, 21 subsets, a 200 × 200 matrix, and a Gaussian 5 mm filter as implemented in the vendor’s software (VG70A, Siemens Healthcare GmbH).
A standard of care static whole-body image was reconstructed by using passes 15–17 of the WB, with ultraHD-PET (PSF + TOF), two iterations, 21 subsets, and a 400 × 400 matrix with a Gaussian 2 mm filter.
[
18F]FDG kinetics were modeled using a two-compartment model based on linear Patlak analysis [
14,
15], as described in detail by A. M. Smith et al. [
16], resulting in the generation of whole-body Patlak slope and Patlak intercept parametric images. Patlak slope, which represents the constant influx rate of [
18F]FDG (Ki
mean, given in mL/(min × 100 mL) = 0.01 × min
−1), was multiplied by the blood glucose level to calculate the metabolic rate of [
18F]FDG (MR-FDG
mean) and is expressed as µmol/(min × 100 mL). Patlak intercept is expressed in percent and represents the distribution volume of free [
18F]FDG (DV-FDG
mean) in the reversible compartments and fractional blood volume [
13]. Semiquantitative measurements were performed in static images using SUV
max, SUV
mean (50% isocontour), and SUV
peak (1 mL sphere).
2.4. Image Evaluation and Segmentation
Parametric images were produced and quantified using syngo.via® 8.2 (Siemens Healthineers, Erlangen, Germany). Volumes of interest (VOIs) were manually delineated in the fused PET/CT images and validated by a certified expert in nuclear medicine with more than five years of experience in PET/CT. VOIs were overlaid on the Ki dataset, DV-FDG, and on the static PET images for data extraction. If necessary, manual coregistration was performed to assure adequate realignment.
2.5. Ground Truth
The final diagnosis was provided by histology, long-time follow-up, and/or as a consensus decision of the institutional interdisciplinary tumor board.
2.6. Statistical Analysis
Differences in the mean values of two groups, features, or methods were tested for significance using the two-sided Student’s t-test. Levene’s test was performed to assess the equality of variance before the t-tests.
One-way ANOVA was performed to compare the dignity (inflammation, benign, or malign) of the different groups for the studied metrics (e.g., DV-FDGmean, MR-FDGmean). An alpha level of 0.05 was used for analysis. Subsequent multiple comparison correction was performed using Tukey’s honestly significant difference procedure. Results of the ANOVA are shown with p values in the main manuscript. Correlation coefficients were calculated according to Pearson and a Pearson correlation coefficient of r > 0.7 was defined as strong, 0.7–0.3 as moderate, and <0.3 as a weak linear correlation. A p-value < 0.05 was considered statistically significant.
The intersection of the false-negative and false-positive rates was defined as the optimal cut-off value. Statistical analysis was performed with SPSS Statistics 28.0 software (IBM Inc., Armonk, NY, USA), MATLAB v. R2022b (The MathWorks, Inc., Natick, MA, USA), and MS Excel 2019 v.2206 (Microsoft corporation, Redmond, WA, USA).
3. Results
3.1. Patient Cohort
Thirty-nine patients met the inclusion criteria for this prospective study between October 2019 and April 2022, of whom 34 consented to study-related dynamic PET acquisition. One patient received further treatment abroad and dropped out of the analysis. Consequently, 33 patients with complete datasets were included in the analysis. Gender distribution was 42% women (14/33) and 58% men (19/33). Male patients were significantly older (68 ± 9 yrs vs. 60 ± 10 yrs, respectively, p = 0.032) and taller (178 ± 9 cm vs. 161 ± 9 cm, respectively, p < 0.001) than female patients with comparable weight (78 ± 22 kg vs. 70 ± 10 kg, respectively, p = 0.053) and BMI (26 ± 6 vs. 27 ± 4, respectively, p = 0.475). The blood glucose level before tracer administration did not differ between the sexes and was 5.44 ± 0.94 mmol/L.
3.2. Pulmonary Lesions
Detailed pulmonary lesion analysis is shown in
Table 1 with 66.7% (22/33) classified as malignant and 33.3% as benign. In one patient, the lung lesion had completely regressed between external CT-scan and PET/CT, so that no lung lesion measurements could be obtained. The final diagnosis was confirmed histologically in 64.6% of the patients (21/33), by follow-up in 21.2% (7/33), and as a consensus decision of the interdisciplinary tumor board in 15.2% (5/33).
3.3. Feasibility of Patlak-PET Data Acquisition
All patients tolerated the complete scheduled acquisition time. No examination had to be discontinued or repeated due to technical difficulties. A representative multiparametric scan is presented in
Figure 2.
3.4. Effect of Quantification Method on Diagnostic Accuracy
Each semiquantitative PET measurement was performed using three different quantification methods: max, mean (50% isocontour), and peak (1 mL sphere). The quantification method showed no significant effect on the AUC, neither for the lung lesions nor for the lymph nodes, as detailed in
Supplementary Tables S1 and S2. For clarity, only the “mean” value is reported in the results.
Malignant lung lesions revealed a significantly higher tumor volume, SUV
mean, Patlak Ki
mean, MR-FDG
mean, and DV-FDG
mean compared with benign lung lesions, as detailed in
Table 2 and
Figure 3. Benign pulmonary nodules were markedly smaller than inflammatory sites, however, this difference was not significant in this cohort (
p = 0.057).
3.5. Lymph Nodes Characteristics
LNM had a significantly higher SUV
mean, Patlak Ki
mean, MR-FDG
mean, and DV-FDG
mean compared to benign and to inflammatory altered LN. Furthermore, LNM presented a significantly larger short- and long-axis diameter compared to benign and to inflammatory-altered LN, as presented in
Table 2. Tumor volume was not a feature that was consistently increased in malignant lesions and could, therefore, not significantly discriminate dignity between the three groups in this cohort.
3.6. Patlak FDG-PET: Dynamic Parameter Evaluation
Liver tissue was chosen as the reference organ and measurements were performed in all patients (n = 33) in tumor-free liver tissue (SUVmean: 2.79; MR-FDGmean: 2.08 µmol/(min × 100 mL); Kimean: 0.406 mL/(min × 100 mL).
Ki
mean and MR-FDG
mean correlated strongly for lung lesions (r = 0.989;
p < 0.001) and LN (r = 0.994;
p < 0.001), so that only MR-FDG
mean is shown in the following figures for reasons of conciseness. Quantified MR-FDG
mean correlated strongly with SUV
mean for lung lesions (r = 0.930;
p < 0.001) as well as LN (r = 0.967;
p < 0.001), as presented in
Figure 4. The correlation between DV-FDG
mean and MR-FDG
mean was slightly lower but still strong and significant (lung lesions: 0.826, LN: 0.760,
p < 0.001).
In distant metastases, MR-FDG
mean quantification showed a strong correlation (r = 0.943;
p < 0.001) with SUV
mean, regardless of the location of metastases or histology of primary tumors, as presented in the scatterplot in
Figure 5A.
When only bone and lung metastases were considered, a strong correlation between SUVmean and Patlak intercept was observed (r = 0.891; p = 0.017).
In contrast, DV-FDG
mean revealed a three-times higher value in an NSCLC liver metastasis (153.63%) compared to the other bone and lung metastases (55.54%), as shown in
Figure 5B. As a result, the correlation with SUV
mean fell below the significance level (r: 0.457,
p = 0.302). However, considering only bone and pulmonary metastases, a strong correlation between SUV
mean and DV-FDG
mean r = 0.891 (
p = 0.017) was found.
3.7. Discriminatory Power between Benign and Malignant Lung Lesions
SUV
mean and the dynamic parameters Patlak Ki
mean, MR-FDG
mean, and DV-FDG
mean revealed very good discriminatory power in the AUC-analysis between benign and malignant lung lesions even at high-significance levels (
p < 0.001), as detailed in
Figure 6 and
Table 3.
MR-FDGmean provided the best discriminatory power between benign and malignant lung lesions with a high AUC of 0.887. At a somewhat lower level, the AUC of DV-FDGmean was 0.818 and that of the SUVmean was 0.827, although the difference did not reach significance in the AUC comparison in this cohort. MR-FDGmean was slightly more specific than SUVmean (81.8% vs. 72.7%, respectively) at a sensitivity of 81.0% (cut-off value of 61.7 µmol/(min × 100 mL)).
Normalizing the SUV
mean of the lung lesions to the SUV
mean of the blood pool in the descending aorta or the hepatic parenchyma did not result in a relevant AUC improvement, as presented in
Table 3.
Regarding CT features, malignant lung lesions presented with significantly larger volume, as detailed in
Table 2. Determination of the pulmonary nodule density was not able to reliably distinguish tumor foci from benign lung lesions (
p = 0.65).
3.8. Discriminatory Power between Benign and Malignant Lymph Nodes
The parametric PET parameters MR-FDG
mean, Patlak Ki
mean, and DV-FDG
mean provided excellent discriminatory power between LNM and benign LN. The AUC of the static PET parameter SUV
mean (AUC 0.993) was slightly, but not significantly, higher than parametric PET parameters, as detailed in the ROC (
Figure 7) and
Table 4. SUV
mean showed the highest sensitivity and specificity within all PET parameters at an optimal cut-off value of SUV 2.6.
For parametric PET, MR-FDG
mean revealed the highest AUC of 0.987 followed by Patlak Ki
mean and DV-FDG
mean with non-significantly lower AUC of 0.958 and 0.948, respectively. Semiautomatic diameter measurements also reached excellent AUC with 0.969 for the short-axis and 0.947 for the long-axis diameter, as shown in
Figure 7 and
Table 4. The calculation of the tumor-to-liver or tumor-to-metastases ratios did not improve AUC for either Patlak Ki
mean, MR-FDG
mean, DV-FDG
mean, or SUV
mean.
3.9. Effect of Distant Metastases on SUVmean, Patlak Kimean, and DV-FDGmean Values of Primary Tumor and LNM
A further analysis was performed to assess the differences in SUV, Patlak Ki, MR-FDGmean, and DV-FDGmean of lung lesions and LNM in patients with or without distant metastasis (M1, contralateral thoracic and/or extrathoracic). LNM presented with significantly higher SUVmean (M1: 13.49 ± 5.65; M0: 3.89 ± 1.89 p = 0.018), Patlak Kimean (M1: 3.09 ± 1.63; M0: 0.63 ± 0.43 mL/min/100 mL, p = 0.031), and MR-FDG (M1: 17.78 ± 9.31; M0: 3.90 ± 1.22 µmol/(min × 100 mL), p = 0.032), but non significantly higher DV-FDGmean (M1: 124.16% ± 44.78; M0: 78.23 ± 25.55, p = 0.129) values in patients with distant metastases (n = 5) compared to M0.
However, primary tumors showed only non-significantly higher SUVmean (10.33 ± 5.37 vs. 5.73 ± 5.37%), Patlak Kimean (3.2 ± 1.85 vs. 1.69 ± 2.1 mL/min/100 mL), MR-FDGmean (18.23 ± 11.01 vs. 9.45 ± 12.60 µmol/(min × 100 mL)), and DV-FDGmean (143.11 ± 91.01 vs. 104.28 ± 101.48%) values in patients with M1 compared to M0.
4. Discussion
This prospective study investigates the additional diagnostic value of whole-body parametric Patlak analysis of [18F]FDG PET in patients with indeterminate lung lesions in a clinical setting. Moreover, we explore the diagnostic performance of dynamic data in the detection of LNM and distant metastases compared to standard static PET scans at 60 min p.i. First, methodologically, we demonstrate the reliability of dynamic whole-body PET/CT acquisition in a multi-bed–multi-timepoint technique with continuous table movement in the clinical routine on a conventional PET scanner. Second, we confirm that the quantified metabolic rate of [18F]FDG (MR-FDG) seems to be at least as accurate in distinguishing malignant from benign findings as the state-of-the-art semiquantitative SUV measurement using 60 min p.i. static scan.
Parametric data from MR-FDG and Patlak Ki correlated strongly with the established SUVmean measurements and had comparable AUCs for the classification of lung lesions. However, a closer look at the ROC indicated a slightly higher specificity in the mid-high sensitivity range for MR-FDG. This finding may indicate that MR-FDG and Ki are slightly more robust than SUV, which is in line with the results of the virtual clinical trial by Ye et al. [
17]. In that study, the Ki was found to be superior to the SUV in the detection of NSCLC and more robust in the case of significant count rate reductions. However, the findings were validated only on a small sample size [
17].
The parametric whole-body dynamic [
18F]FDG PET measurements of our trial were consistent with the limited data available from previous studies [
18]. In direct comparison to single-bed dynamic PET measurements published by Yang et al., our results demonstrate slightly higher SUVs in the primary tumor (M0: SUV
mean 5.73 vs. 5.23; M1: 10.33 vs. 8.41), and considerably lower Ki values (M0: 0.0169 min
−1 vs. 0.026; M1: 0.032 min
−1 vs. 0.050) [
6]. Similar results were also found for LNM, whose uptake was also shown to be dependent on the presence of distant metastases (SUV
mean: M0: 3.89 vs. 4.22; M1: 13.49 vs. 5.57) [
6].
While SUV
mean measurements are generally accepted in the clinical setting, the use of Ki
mean is not validated yet. Here, the MR-FDG values of the lung tumors differed up to a factor of two compared to the dynamic single-bed measurements at comparable SUV
mean. This effect was more emphasized and indeed dependent on the presence of distant metastases (Patlak Ki
mean: M0: 0.0063 vs. 0.016 min
−1, M1: 0.031 vs. 0.033 min
−1) [
6]. Notably, our data showed a significantly stronger correlation between SUV
mean and Patlak Ki
mean (r: 0.93–0.97 vs. 0.76–0.88) compared to the data published by Yang et al. [
6]. Such varying strength of correlation between two parameters, which were calculated at one site each, indicate that the Ki values may depend on the calculation method. However, this must be further investigated.
In addition, it is also important to consider that although the magnitude increments of SUV
mean and Patlak Ki
mean or MR-FDG
mean are quite similar, they represent different physiological information. SUV
mean is the sum of metabolized [
18F]FDG-6P trapped in the compartment and un-metabolized [
18F]FDG, while MR-FDG solely reflects metabolized [
18F]FDG-6P activity [
18].
Furthermore, data on our DV-FDG measurements, which represents the combined distribution volume of free [
18F]FDG in blood and tissue (reversible compartment), also revealed strong correlations with trapped [
18F]FDG measured within MR-FDG and Patlak Ki
mean (irreversible compartment) [
18]. Interestingly, the only hepatic metastasis in our cohort was visually more distinct and focal in the parametric DV-FDG image, compared to the other parametric parameters. Furthermore, this lesion presented with a remarkably higher DV-FDG value, when compared to the lung or bone metastases. One potential explanation for this effect in the liver metastasis is a previously reported increment of dephosphorylation of the trapped [
18F]FDG-6P in liver tissue [
18]. High dephosphorylation activity would result in less irreversible trapping and significant efflux of the initially trapped [
18F]FDG-6P via the bidirectional GLUT (esp. GLUT 1) transporter out of the cell and back into plasma [
18]. This would result in higher DV-FDG values since the reversible compartment also includes both free [
18F]FDG in blood and tissue as well as some [
18F]FDG-6P [
18]. Even if the value of DV-FDG has caused some controversy [
19], our data are supportive of investigations evaluating DV-FDG as a potential imaging biomarker for liver metastases.
Interestingly, in our cohort, the diagnostic performance of Patlak Ki
mean and MR-FDG seems to achieve at least equal or higher discriminatory power in the detection of mediastinal LNM when compared to the dual-time-point (DTP) dynamic PET using an SUV retention index (RI-SUV) between 1 h and 2 h p.i. by Shinya et al. [
9] or the DTP data presented in the largest meta-analysis by Shen et al. [
7] (AUC 0.958 vs. 0.794 and 0.9331) on lesion-based analysis. In detail, our MR-FDG
mean quantifications presented with higher sensitivity of 92% vs. 74% at a defined specificity of 76% and higher specificity of 89% vs. 76% at a defined sensitivity of 74% compared to the DTP-based RI-SUV estimation published by Shinya et al. [
9].
Regarding the performance of dynamic parameters for the detection of distant metastases, there are still insufficient data in the literature. The parametric [18F]FDG dynamic data presented in this study, however, provide the largest published cohort with histologic validation. MR-FDG was shown to be a robust parameter with a very strong correlation to SUVmean regardless of the histology of the primary tumor or location of metastasis (bone, lung, or liver).
Limitations
There are several limitations in this prospective pilot study. First, the sample size of LNM and distant metastases is relatively small, even though it represents one of the largest published collectives. However, due to large effect sizes, the data presented are significant and, therefore, might enable a pre-conclusive analysis.
In addition, some of the lesions could not be confirmed by biopsy; thus, the diagnosis had to be confirmed based on the conclusion of the interdisciplinary tumor board, as is the gold standard for many lesions.
Data acquisition was performed within a single-center study setting; thus, the intercomparability of measurements between different PET scanners cannot be evaluated.