In this retrospective analysis datasets of 43 patients who underwent initial F-18-FDG PET/CT for staging of suspected lung cancer at our facility between December 2018 and December 2020 were included. Patients with prior resection of the primary tumor were excluded. The study design was approved by the local ethics committee of the University of Münster (AZ 2019-024-f-S, 2021-172-f-S), and was performed in accordance with the 1964 Helsinki declaration and its later amendments. The need for written informed consent was waived due to the retrospective nature of the study.

The patients fasted overnight before the PET/CT scan. They received 3 MBq/kg body mass of F-18-FDG i.v. approximately one hour prior to the scan which was performed on a Biograph mCT (Siemens Healthcare GmbH, Erlangen, Germany) capable of time-of-flight and continuous bed motion (axial PET field-of-view, 21.8 cm; spatial resolution at center, 4 mm full width at half maximum; sinogram sizes, 400 × 168; time-of-flight bins, 13) [25]. Patients were scanned in a supine position with the arms above the head. During the examination, the respiratory gating system AZ-733 V (Anzai Co., Tokyo, Japan) recorded respiratory signals for subsequent gating (belt gating, BG) and motion correction.

Scanning ranges were from the head or neck down to the proximal femur. End-expiratory low-dose CT scans were performed (tube voltage, 120 kV; effective current, 18 mAs; slice thickness, 3.0 mm; duration, 10–20 s) followed by PET in continuous bed motion (free breathing; speed, 1.1 mm/s; duration, 500–900 s).

Three different PET reconstructions were investigated within this study (Fig. 1): (1) Static reconstruction without motion correction (“static”); (2) elastic motion-corrected reconstruction based on the belt gating signal (“BG-MC”); and (3) elastic motion-corrected reconstruction based on PET raw data-driven gating signal (“DDG-MC”).

Reconstruction workflow used for the three PET images (“static,” “BG-MC” and “DDG-MC”) performed within this study

The applied DDG algorithm is based on a spectral analysis of continuous bed motion PET raw data and is described in detail elsewhere [17, 21]. Briefly, it divides the raw data into axial regions of 80 mm length, where measured events are back-projected into the most likely origin voxel according to their time-of-flight bin. The predominant respiratory frequency was then identified by the maximum in the power spectrum of the standard deviation along the anterior–posterior axis over time. Voxels that demonstrated fluctuations close to this frequency were then used to define a mask of regions affected by respiration. Respiratory signals for each axial region were then calculated by phase- and mask-weighted summation of voxel time–activity curves and finally concatenated and normalized to give an overall DDG signal for the whole PET scan.

Signals from both sources were used for elastic motion-corrected PET reconstructions by first reconstructing the “optimal gate” comprising coincidence data from the narrowest signal amplitude interval covering 35% of the total data, giving a good compromise between motion resolution and data statistics, and then using mass-preserving optical flow techniques to determine a motion vector field between the gated and a static reconstruction. This vector field was then finally used in an effective deblurring step within a motion-corrected image reconstruction [18, 19], resulting in BG-MC and DDG-MC datasets.

All reconstructions were based on an ordinary Poisson ordered subset expectation maximization (2 iterations, 21 subsets, 2 mm full width at half maximum Gaussian post-reconstruction filter, 400 × 400 image matrix, 2.04 × 2.04 × 2.03 mm3 voxel volume; e7 toolbox, Siemens Healthcare GmbH, Erlangen, Germany) with point-spread-function and time-of-flight data, normalization, and random correction; attenuation and scatter correction were based on the measured CT data. Overall, three PET and one CT image dataset per patient were thus subsequently analyzed.

All PET and CT images were anonymized and sent to a syngo.via workstation (Oncology tool, Siemens Healthcare GmbH, Erlangen, Germany) where they were presented independently to two nuclear medicine specialists (BN, WR) with more than five years of experience in PET/CT imaging. One of the three PET reconstructions, the CT image and a fused PET-CT image were made available to a reader. The three different PET reconstructions (static, BG-MC, DDG-MC) for any given scan were presented in random order and in different sessions in an interval of at least 2 weeks to reduce bias. The readers were blinded for the actual type of reconstruction.

The lymph node (N) and distant metastasis (M) status was assessed, with the N rating further divided into the three different lymph node regions N1 (ipsilateral peribronchial and/or hilar lymph nodes), N2 (ipsilateral mediastinal and/or subcarinal lymph nodes), and N3 (contralateral mediastinal and/or hilar, as well as any supraclavicular lymph nodes), following the TNM staging system for lung cancer of the American Joint Commission of Cancer (AJCC) and the Union Internationale Contre la Cancer (UICC) [26]. For every reconstruction, these three N regions and the M status were independently rated on an ordinal scale \(s\) ranging from 1 (“certainly negative”), 2 (“probably negative”), 3 (“doubtfully negative”), 4 (“doubtfully positive”), 5 (“probably positive”), to 6 (“certainly positive”). Derived from this score, a simplified dichotomous score \(d\) was defined as 0 for negative findings (scale values of 1, 2, 3) and 1 for positive findings (scale values of 4, 5, 6).

Finally, to quantify the subjective certainty of the readers, an ordinal certainty score was calculated as

$$c = \left| \right| + 0.5$$

with 1 denoting least certainty and 3 denoting highest certainty.

Additionally, the primary tumor and the most prominent lymph nodes visible in each region N1, N2 and N3 were characterized by their standardized uptake values SUVmax, and SUVmean, and the metabolic tumor volume (MTV) in each reconstruction.

Analyses were performed using R statistical software version 3.6.1 (The R Foundation, r-project.org). All reported p values are two-sided. Normally distributed data were described using mean and standard deviation. Non-normally distributed data were described using median and interquartile range. Normality was assessed by analysis of histograms and skewness statistics.

Interobserver agreement for TNM staging using the ordinal scale \(s\) was assessed using Cohen’s weighted kappa statistics. In the primary statistical analysis differences in the ordinal score values \(s\) between reconstruction methods were assessed for each region by nonparametric analysis of longitudinal data in factorial experiments using the R package nparLD [27], as were differences in the certainty score \(c\). The method accounts for dependencies between measurements on the same patient (i.e., for a given region each patient provides a measurement per reconstruction method and reader, resulting in six observations per patient). A multiple comparison procedure based on the closed testing principle [28] was applied to each region using a (multiple) significance level of 0.05 per region. Following this principle, a single pairwise comparison was considered significant, if both the overall comparison and the pairwise comparison resulted in a p value ≤ 0.05.

SUV and MTV showed a non-normal distribution in histograms analysis. Differences in SUV and volumes between methods were assessed in an exploratory analysis using Friedman’s test. Wilcoxon signed-rank tests were applied as post hoc procedure. p values ≤ 0.05 were considered significant.