Thirteen patients with NSCLC from two clinical trials carried out at the Amsterdam University Medical Center, were included in the present study. The clinical trials were approved by the ethical review board of the Amsterdam University Medical Center and informed consent was obtained from all patients. The ATTAIN trial was registered under clinicaltrials.gov number NCT05157659. The SHARP trial was registered under clinicaltrials.gov number NCT05701176. The details of both trials are described in supplemental material SM1.

Ten patients with early stage NSCLC with a resectable tumour were enrolled in the ATTAIN trial. PET imaging was acquired using an Ingenuity TF PET/CT scanner (Philips Healthcare). The axial field of view (FOV) was positioned over the thorax containing the primary tumour. A low dose CT (ldCT) scan was obtained for attenuation correction followed by a dynamic [18F]F-AraG PET scan. For the first five patients, the duration of the dynamic PET scan was 90 min. Initial analyses showed that the scan time could be reduced, hence, for all other patients, the scan duration was 70 min. One patient did not complete the full 70 min and was scanned 60 min. Six patients underwent a second scan within four days after the first dynamic PET scan. During scanning, blood samples were manually drawn at 5, 10, 15, 25, 40, 60 and 90 min post injection (p.i.). For seven patients, both arterial and venous sampling was obtained for correlation purposes. From the other three patients, the full sampling protocol could not be performed, so either arterial or venous sampling was collected.

Three patients with advanced stage NSCLC enrolled in the SHARP trial. PET imaging was acquired using a 106 cm long axial FOV Biograph Vision Quadra PET/CT scanner (Siemens Healthineers). A ldCT scan was obtained for attenuation correction followed by a 70 min dynamic [18F]F-AraG PET scan. During scanning, venous blood samples were manually drawn at 5, 10, 15, 25, 40 and 60 min p.i.

In both trials, PET scans were corrected for detector normalization, decay, dead time, attenuation, randoms and scatter. PET scans from the Ingenuity TF PET/CT were reconstructed using 3D-RAMLA and had a voxel size of 4 × 4 × 4 mm3 with a spatial resolution of 4.6–5.7 mm in full width at half maximum (FWHM). PET scans from the Biograph Vision Quadra PET/CT were reconstructed using PSF-TOF + 4 mm Gauss (EARL2 [12, 13]) and had a voxel size of 3.3 × 3.3 × 2 mm3 with a spatial resolution of 3.3–4.6 mm in FWHM. Depending on the respective scan times, all scans were reconstructed into 19, 20 or 22 frames (1 × 15, 3 × 5, 3 × 10, 4 × 60, 2 × 150, 2 × 300, and 4, 5 or 7 × 600 s).

For each sample, radioactivity concentrations (AC) in plasma and whole-blood were determined using a γ-counter. Furthermore, plasma was analysed for radiolabelled metabolites of [18F]F-AraG by solid-phase extraction (SPE). In brief, 1 mL of plasma was diluted with 2 mL water and loaded onto an activated Sep-Pak tC18 cartridge (Waters, Milford, MA, USA, 6 cc, 1 g). Vacuum was applied and the diluted plasma was passed over the SPE column, whereafter the cartridge was washed with 5 mL of water. The column was subsequently eluted with 5 mL methanol. Radioactivity in all three fractions (plasma, water, methanol) was measured. The first two fractions represent polar radiolabelled metabolites of [18F]F-AraG and the third intact [18F]F-AraG.

[18F]F-AraG PET scans were visually assessed by a nuclear medicine physician for uptake in the primary tumour, metastases, lymph nodes and healthy organs. The corresponding volumes of interest (VOI) were manually delineated on the dynamic [18F]F-AraG PET, using the ACCURATE tool [14]. Both anatomical information from the ldCT scan and uptake on the PET scans were used to locate the VOI. In case of a small mismatch between ldCT and PET because of breathing motion, delineations were corrected by delineating only the overlap between ldCT and PET. Mean lesion AC over time (time-activity curves (TACs)) were obtained for each VOI for the duration of the dynamic PET scan.

A region of interest with a diameter of 17 to 20 mm was placed within the lumen of the ascending aorta on 5 to 8 consecutive axial slices (volumes ranging from 4.2 to 6.7 mL) to derive the blood AC from the images. These data will be used to generate an image derived input function (IDIFs).

Organs of interest within the FOV were manually delineated on the dynamic [18F]F-AraG PET scan using the ACCURATE tool [14]. Delineations of the lungs were obtained using a standard ldCT scan threshold. The liver, spleen, kidneys and muscle tissue (i.e., the biceps and triceps of both arms) were manually delineated based on the anatomical location on the ldCT scans, for the areas of the organs that were within the FOV. The delineations were corrected based on uptake on the [18F]F-AraG PET scan. The myocardium and thyroid glands were delineated based on uptake on the [18F]F-AraG PET scan. The bone marrow was delineated by placing 3 fixed size VOIs within in 3 separate thoracic vertebrae. SUVBW was obtained for each organ for the duration of the dynamic PET scan to evaluate the biodistribution of [18F]F-AraG. All delineations were adjusted to minimize spill-over from adjacent blood vessels using the first timeframes (40–60 s p.i.) of the dynamic scan.

All venous samples were corrected for the bias between venous and arterial sampling (see supplemental Fig. S1 and S2) by applying a correction factor to the whole-blood and plasma samples for each time point. For the ATTAIN trial, acquired using the Ingenuity TF PET/CT scanner, quantification of the IDIF showed misalignment in the shape of the curve when compared to whole-blood samples. This appeared to result from low signal-to-noise ratios on the PET at later time points due to scatter correction issues as assessed with phantom experiments (data not shown). Consequently, high noise levels were present in the aorta due to surrounding high-uptake areas (i.e., the myocardium and liver), which only influenced the IDIF at later time points with low AC in the ascending aorta (~ 700 Bq/mL). To correct for this, a bi-exponential fit through the whole-blood samples (from 5 min p.i. until the end of the scan) was combined with the first five minutes of the IDIF (not influenced by scatter correction issues) to obtain a whole-blood input functions. For the SHARP trial, acquired using the Biograph Vision Quadra PET/CT scanner, this issue did not appear. A single exponential function fitted to the samples was compared to the IDIF-tail (> 800s) to obtain a calibration factor for the IDIF. This calibration factor was derived by minimizing the difference in IDIF and sample activity concentration values. The IDIFs were scaled with this calibration factor (towards the samples) to obtain a whole-blood input function. Subsequently, all whole-blood input functions were corrected for plasma/whole-blood ratios, which were extrapolated to all time points, and parent fractions, through which a Hill function was fitted. For two patients (i.e. ATT03 and ATT04), assessment of the parent fraction was unsuccessful due to low extraction efficiency of the SPE material. The method was optimised for the following patients. For these two patients, the median parent fraction of the other patients was used for correcting the input functions. Metabolite corrected plasma input functions were used for pharmacokinetic modelling.

Pharmacokinetic modelling was performed on the lesion TACs using in-house developed software build in MATLAB (MATLAB version 7.04, The MathWorks, Inc., Natick, Massachusetts, United States). The tool applies weighted non-linear regression with constraints on the pharmacokinetic rate constants [15]. A single-tissue reversible model (1T2k), a two-tissue irreversible model (2T3k) and a two-tissue reversible model (2T4k) were fitted to the TACs with an additional fit parameter for the blood volume fraction (Vb) to assess the preferred pharmacokinetic model [11]. The following pharmacokinetic rate constants were derived from the analyses; K1 [min− 1] representing the [18F]F-AraG influx from plasma to tissue, k2 [min− 1] representing the [18F]F-AraG efflux from tissue to plasma, k3 [min− 1] representing the rate [18F]F-AraG uptake to activated T-cells and k4 [min− 1] representing the release rate of [18F]F-AraG from activated T-cells. The rate constant Ki, representing the net influx rate, and the VT, representing the distribution volume, were calculated from the obtained pharmacokinetic rate constants [11]:

$$\:_=\frac_*_}_+_}\quad\quad_=\frac_}_}*(1+\frac_}_})$$

The SUVs corrected for body weight (SUVBW), body surface area (SUVBSA), and lean body mass (SUVLBM) were derived for the following time intervals: 20–30, 30–40, 40–50, 50–60 and 60–70 min p.i. Tumour-to-blood ratio (TBR) and tumour-to-plasma ratio (TPR) were derived for the following time intervals: 20–30, 40–50 and 60–70 min p.i.

The preferred kinetic model was selected based on the goodness of fit, both visually assessed and tested with the Akaike information criterion (AIC) [16]. Since some tumour lesions were included for both test and re-test, and arterial and venous sampling, only one data point per lesion was included for correlation analyses. If available, the scan corresponding with the arterial sampling was selected. Pearson correlation was performed to assess the association between simplified uptake measures and the outcome parameter of the optimal pharmacokinetic model. The significance level (p-value) was set at 0.05. Statistical analyses were performed using R Statistical Software (v3.6.1; R Core Team, 2019).