Parametric analysis of dynamic long axial field-of-view 18F-FDG PET/CT studies Authors: Charlotte L. C. Smith, Gerben J. C. Zwezerijnen, Marijke E. den Hollander, Henricus N. J. M. Greuter, Nienke R. Gerards, Josée Zijlstra, C. Willemien Menke-van der Houven van Oordt, Idris Bahce, Maqsood Yaqub, Ronald Boellaard Published: 2026 ABSTRACT BackgroundLarge-axial field of view (LAFOV) PET/CT systems have improved whole-body dynamic 18F-FDG imaging, enabling improved quantitative assessment of tracer kinetics. While non-linear regression (NLR) models remain the most unbiased approach, voxel-wise implementation can suffer from convergence problems, resulting in increased uncertainties. Parametric methods such as Patlak, Generalized Patlak (gPatlak), basis function method irreversible (BFM-I), and basis function method reversible (BFM-R) offer alternatives for estimating parametric Ki and, if applicable, K1 values. This study aims to determine the most suitable parametric approach for quantifying 18F-FDG kinetics in various healthy organs and tumour lesions using a LAFOV PET/CT system.MethodsThirteen patients with malignancies underwent 70-minute dynamic 18F-FDG PET/CT scans on a LAFOV PET/CT system after receiving an intravenous bolus injection of 3.0 MBq/kg. First, the data were split on preference for irreversible (2T3k) or reversible (2T4k) kinetics. Ki from Patlak (t* = 30 min) and BFM-I (t* = 0 min) were compared with those derived from the 2T3k model. Ki from gPatlak (t* = 10 min) and BFM-R (t* = 0 min) were compared with those obtained with 2T4k. Ki and K1 values were extracted from the brain, liver, spleen, lung, kidney, upper leg muscle, and up to three tumour lesions per patient.ResultsThe liver, myocardium, lungs, and muscle favoured the irreversible 2T3k model, while preferences for the kidney, brain, spleen, and tumour lesions varied across patients. For organs preferring irreversible kinetics, BFM-I (R = 0.99) outperformed Patlak (R = 0.97) in Ki agreement with the 2T3k model. For reversible kinetics, BFM-R (R = 0.95) showed better correspondence compared to gPatlak (R = 0.90) with the 2T4k model. K1 obtained by BFM-I and BFM-R also strongly correlated with those derived from compartment models.ConclusionsLAFOV PET/CT systems enable accurate whole-body kinetic modelling of 18F-FDG uptake. BFM-I and BFM-R demonstrated the highest agreement with the compartment models and are recommended when full dynamic PET scans are feasible, offering accurate Ki as well as K1 estimations across healthy organs and tumour lesions.Supplementary InformationThe online version contains supplementary material available at 10.1186/s40658-026-00859-3. Background Large axial field-of-view (LAFOV) PET/CT systems significantly enhance whole-body dynamic 18F-FDG PET imaging, enabling a comprehensive quantitative assessment of glucose metabolism in key organs [1]. Dynamic PET imaging with a LAFOV PET/CT system provides insight into tracer uptake and binding across all key organs, enabling whole-body kinetic modelling [1]. When performed accurately, kinetic modelling improves diagnostic accuracy, response prediction, and treatment planning [2]. The reference method for modelling 18F-FDG kinetic parameters is based on fitting a two-tissue compartment model (2TC model) to the observed time activity curves (TACs). However, for voxel-wise parametric imaging, complex models such as a 2TC model may not be ideal, as they are slow and prone to fitting convergence problems at voxel-by-voxel noise levels [3]. Therefore, simpler parametric methods that require fewer parameters and assumptions may be better suited for parametric imaging, as they are more robust against noise. However, selecting an appropriate parametric method remains essential to ensure that different tissue kinetics are captured accurately without introducing bias into the parameter estimates [4]. The most commonly used parametric method for 18F-FDG is Patlak linearization, which assumes irreversible trapping of FDG in all tissues to estimate the net influx rate (Ki). However, several studies have reported that some tissues, such as the brain, liver, lungs, and certain tumours, might show reversible kinetics [3, 5–9]. A previous study of Hu et al. [10] already looked into the correlation between voxel-wise compartment modelling with Patlak analysis, but they did not study other methods. They, together with other previously mentioned studies, demonstrated that the use of an irreversible model when tissue shows reversible kinetics can lead to underestimation of Ki [3, 5–8, 10]. To address this, Karakatsanis et al. [6] proposed Generalized Patlak (gPatlak) for tissues showing reversible kinetics as gPatlak incorporates an additional kinetic parameter to account for the reversible compartment [6, 11]. Another promising approach is the basis function method (BFM). Until now, BFM for dynamic 18F-FDG PET scans has only been studied in the brain [3], but it may also be suitable for whole-body dynamic 18F-FDG PET imaging. BFM introduces less variability than the 2TC model fits using non-linear regression (NLR) while accounting for blood fraction (VB), which is not taken into account with Patlak and gPatlak analyses. Therefore, BFM not only estimates Ki but also K1 (rate constant from arterial blood to tissue: K1=extraction * flow) [12]. BFM is a computationally efficient algorithm using a set of predefined basis functions (BF) to analyse for the best parametric fit and can be implemented for both irreversible (BFM-I) and reversible (BFM-R) kinetics [12]. Accurate selection of the appropriate parametric method is essential for reliable estimation of kinetic parameters and valid clinical interpretation. This study aims to determine suitable parametric approaches for quantifying 18F-FDG kinetics in various healthy organs and tumour lesions using a LAFOV PET/CT system. First, we identified whether a reversible or irreversible 2TC model provided the best fit on the organ level using the Akaike Information Criterion (AIC). Thereafter, we compared kinetic parameters derived from these NLR model fits with organ averaged parameters obtained from Patlak and BFM-I for tissues best described by irreversible kinetics, and with gPatlak and BFM-R for tissues showing reversible kinetics. Methods Study population Thirteen patients were included from January 2022 to March 2024 at Amsterdam UMC, location VUmc. All patients had a clinical indication for an 18F-FDG PET/CT scan for staging of various malignancies. All patients underwent a 70-min dynamic 18F-FDG PET scan on the Siemens Biograph Vision Quadra PET/CT system (Siemens Healthineers, Knoxville, TN, USA) (LAFOV PET/CT system) at Amsterdam UMC, location VUmc. Written informed consent was obtained from all patients, and the study was approved by the Medical Ethics Review Committee of the Amsterdam UMC (NL80924.029.2). Imaging protocol The scans were performed according to the European Association of Nuclear Medicine Research Ltd 2.0 guidelines [13]. Patients received 3.0 MBq/kg 18F-FDG at the start of the dynamic acquisition through an intravenous bolus injection and were fasting for 4–6 h before tracer administration. Plasma glucose levels were < 7.0 mmol/L [13]. The scans were acquired in one bed position covering 106 cm, including the skull vertex to mid-thigh. The 70-min dynamic PET scans were binned into 20 frames (1 × 15 s, 3 × 5 s, 3 × 10 s, 4 × 60 s, 2 × 150 s, 2 × 300 s, 5 × 600 s). Reconstructions adhered to the EARL 18F-FDG standard 2 (EARL2) reconstruction protocol, consisting of the PSF + TOF OSEM algorithm with 5 subsets, 4 iterations, matrix size of 220 × 220, voxel size of 3.2 × 3.2 × 2 mm3, and 4 mm Gaussian filter [13]. Imaging analysis 18F-FDG PET/CT scans were analysed with the in-house build semi-automated analysis software tool ACCURATE (developed in IDL version 8.4 (Harris Geospatial Solutions, Bloomfield, USA)) [14]. The image-derived input function (IDIF) was obtained from whole-blood 18F-FDG activity from the PET image, calculated by placing ten consecutive circular regions of interest (ROI) in the lumen of the ascending aorta (AA) [15]. The unaffected liver, brain (grey matter), lungs (caudal right and cranial right), spleen, muscle (upper right leg), kidney, myocardium, and up to three tumour lesions per patient were delineated by manually placing a volume of interest (VOI) in the tissue or by semi-automated segmentations. Up to 3 tumour lesions per patient were selected to ensure a proper balancing of the number of data points between the patient studies. The brain (grey matter), kidney, myocardium, and lesions were delineated semi-automatically based on 18F-FDG uptake on the dynamic PET scan. The semi-automatic delineations were delineated based on the last frame (60–70 min post-injection). The liver, lungs, spleen, and muscle were delineated by manually placing a VOI in the organs. For all delineations, mean activity concentration over time (TAC) was obtained for the duration of the dynamic 18F-FDG PET scan. The same VOIs were used when extracting average Ki and/or K1 values from the parametric images. VOI sizes and segmentation methods are summarized in Table 1. Table 1Volume-of-interest sizes for the healthy organsRegionVOI/ROI sizeVolume VOI/ROI (mL)Ascending aorta10 times 1.0 cm ROI1.9602Liver3.0 mL spherical VOI14.098Spleen1.5 mL spherical VOI1.753Muscle (right upper leg)2.0 mL spherical VOI4.173Lung caudal (right)1.5 mL spherical VOI1.753Lung Cranial (right)1.5 mL spherical VOI1.753RegionSegmentation methodBrain region (grey matter)MV3KidneySUV25MyocardiumSUV40TumourMV3MV3, majority vote 3, ROI, region of interest; SUV25, fixed SUV threshold of 2.5 g/mL; SUV40, fixed SUV threshold of 4.0 g/mL; VOI, volume of interest. Table 1Volume-of-interest sizes for the healthy organsRegionVOI/ROI sizeVolume VOI/ROI (mL)Ascending aorta10 times 1.0 cm ROI1.9602Liver3.0 mL spherical VOI14.098Spleen1.5 mL spherical VOI1.753Muscle (right upper leg)2.0 mL spherical VOI4.173Lung caudal (right)1.5 mL spherical VOI1.753Lung Cranial (right)1.5 mL spherical VOI1.753RegionSegmentation methodBrain region (grey matter)MV3KidneySUV25MyocardiumSUV40TumourMV3MV3, majority vote 3, ROI, region of interest; SUV25, fixed SUV threshold of 2.5 g/mL; SUV40, fixed SUV threshold of 4.0 g/mL; VOI, volume of interest. Volume-of-interest sizes for the healthy organs MV3, majority vote 3, ROI, region of interest; SUV25, fixed SUV threshold of 2.5 g/mL; SUV40, fixed SUV threshold of 4.0 g/mL; VOI, volume of interest. Kinetic modelling Pharmacokinetic modelling was performed with the custom-built in-house software developed in MATLAB (MATLAB version 7.04, The MathWorks, Inc., Natick, Massachusetts, United States). IDIFs were obtained from the AA, and TACs for healthy organs and tumour lesions were obtained using predefined VOIs. A region-specific input delay correction was obtained for all TACs. Subsequently, NLR model analysis with constraints on the pharmacokinetic rate constants was performed [16]. The NLR fits were visually inspected to confirm that the model accurately represented the measured data. A summary of all applied constraints is provided in Supplemental Table 1. A two-tissue irreversible model (2T3k) and a two-tissue reversible model (2T4k) were fitted to the TACs and the following pharmacokinetic rate constants were derived: K1 [min− 1] representing the 18F-FDG influx from blood to tissue, k2 [min− 1] representing the 18F-FDG efflux from tissue to blood, and k3 [min− 1] representing phosphorylation [17]. For the 2T4k model, a k4 [min− 1] representing dephosphorylation was also derived as a pharmacokinetic rate constant. Fractional blood volume, VB, was also added as a fit parameter for both the 2T3k and 2T4k pharmacokinetic models [17]. The net influx rate, Ki [mL·cm− 3·min− 1], was calculated from the obtained pharmacokinetic rate constants [17]. For both pharmacokinetic models, Ki was calculated as follows: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{K}_{i}=\:\frac{{K}_{1}*{k}_{3}}{{k}_{2}+{k}_{3}}$$\end{document} Parametric analysis Ki and if applicable K1 were obtained for all voxel TACs by Patlak linearization (using t*= 30 min) and gPatlak (using t*= 10 min) where t* indicates the start time for the data used in Patlak for the linear fit. BF calculation was based on the following equation: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:BF=\:{C}_{Plasma}\otimes\:\:{e}^{-\theta\:*t}$$\end{document}. For BFM-I, 50 BFs were used, of which one was an integral of the IDIF to account for irreversible trapping. Theta ranged from 0.1 to 2.2 min− 1. For BFM-R, two sets of predefined BFs were used as the tissue impulse response function was described by two exponential functions. The first set contained 50 BFs with theta ranging from 0.1 to 2.2 min− 1. The second set contained 20 BFs with theta ranging from 0.00001 to 0.1 min− 1 [3, 12, 18]. While fitting the model, each combination of BFs for each of the two sets was evaluated, and therefore resulting in a total of 1000 combinations. Statistical analysis Statistical analyses and data visualization were performed in R (version 4.4.0). For each organ in each patient, the preferred kinetic model (2T3k or 2T4k) was determined based on AIC [19]. Additionally, the VOIs preferring irreversible kinetics were compared to the parametric models based on irreversible assumptions, and VOIs preferring reversible kinetics were compared to the parametric models based on reversible assumptions to decrease potential bias. This means that the kinetic parameters obtained by Patlak and BFM-I were compared to the organs best described by the 2T3k model, and kinetic parameters derived by gPatlak and BFM-R were compared to the organs best described by the 2T4k model. These comparisons were performed on an individual-organ basis for all patients. The relationship in kinetic parameters of the parametric models (Patlak, gPatlak, BFM-I, or BFM-R) and those derived from the reference NLR 2TC model fits (2T3k or 2T4k) was visualized using scatter plots and summarized with regression lines. Statistically significant associations (p ≤ 0.050) were analysed with a Pearson correlation or Spearman’s rank-order correlation test, when appropriate. Results Patient demographics Thirteen patients (38.5% female) with various malignancies were included in the study. The cohort included one patient with Hodgkin lymphoma, four with multiple myeloma, one with breast cancer, one with melanoma, one with a Kaposi sarcoma, two with diffuse large B-cell lymphoma (DLBCL), one with a T-cell lymphoma, one with a mantel cell lymphoma, and one with a salivary gland tumour. None had a history of type I or II diabetes mellitus. Patient demographics are summarized in Table 2. Eleven patients exhibited 18F-FDG uptake in at least one tumour, and a total of 27 tumour lesions were selected. One tumour lesion was excluded as we could not accurately derive the 2TC model fits. This was probably due to movement of the tumour lesion, as it was a small lesion located near the chest wall and heart. Model selection based for the different tumour lesions based on AIC is illustrated in Supplemental Table 2. The myocardium could only be derived in nine of the thirteen patients, as the myocardium of the other four patients did not show sufficient 18F-FDG uptake. Table 2Patient demographicsCharacteristicsValueGender (F/M), number5/8Age (years), mean (range)60.9 (48–79)Height (cm), mean (SD)176.7 (10.8)Weight (kg), mean (SD)78.0 (19.2)Net injected dose (MBq), mean (SD)240.5 (59.9)cm, centimetre; F, female; kg, kilogram; M, male; MBq, mega Becquerel; SD, standard deviation. Patient demographics cm, centimetre; F, female; kg, kilogram; M, male; MBq, mega Becquerel; SD, standard deviation. Model preference Based on the AIC, each organ in each patient was classified as preferring either irreversible or reversible kinetics. Among the 26 tumour lesions evaluated, 14 showed a preference for the irreversible 2T3k model. For the healthy organs, the liver, myocardium, and lungs consistently showed a preference for the 2T3k model across all patients. The muscle predominantly favoured the 2T3k model, with 11 out of 13 patients showing a preference for the irreversible pharmacokinetic model. In contrast, model preference for the kidney, brain, and spleen varied more between patients. The brain showed a preference for irreversible kinetics in 7 patients, the kidney in 9 patients, and the spleen in 7 patients. Model preference across all organs and patients is summarized in Fig. 1. Fig. 1Model preference for the two-tissue irreversible model (2T3k) or two-tissue reversible model (2T4k) based on the Akaike Information Criterion (AIC) Model preference for the two-tissue irreversible model (2T3k) or two-tissue reversible model (2T4k) based on the Akaike Information Criterion (AIC) NLR 2T3k model preference with Patlak & BFM-I Figure 2 illustrates the Ki parametric maps from a representative patient, obtained using Patlak (a) and BFM-I (b). The parametric maps of Patlak and BFM-I maps appeared nearly identical, except for the increased uncertainty in the aortic region for BFM-I. This is likely due to difficulties in separating tissue and blood signals when VB was above 50%, resulting in unrealistically high Ki values. Therefore, we also added the Ki parametric map obtained by BFM-I, where we removed voxels with a VB larger than 50% (c). This resulted in a cleaner image compared to the BFM-I images without removed voxels, and the voxels removed align with large blood pool structures All healthy organs and tumour lesions showing a preference for the 2T3k model, based on AIC, were compared to Patlak linearization and BFM-I. In one patient, the myocardium exhibited a markedly higher Ki value using BFM-I compared to the other patients. The myocardium can be difficult to quantify due to continuous cardiac and respiratory motion degrading the signal at various levels. This data point was excluded from the regression fits and correlation coefficients. Figure 3a and b compare the Ki values obtained from the Patlak and BFM-I models with the Ki values derived by the 2T3k pharmacokinetic model. A strong correlation was observed between Ki values obtained by Patlak and the 2T3k model. However, Patlak systematically underestimated Ki across all tissues, as Patlak does not account for VB. BFM-I demonstrated a strong correlation and agreement with the 2T3k pharmacokinetic model, with a regression line nearly identical to the line of identity. All correlation coefficients and regression parameters are summarized in Table 3 To examine the impact of applying a reversible parametric model to tissue preferring irreversible kinetics, we compared Ki estimates from gPatlak and BFM-R to those obtained by the 2T3k model. The correlation and regression fit coefficient of these analysis, based on healthy organs and tumour lesions, are illustrated in Supplemental Fig. 1 and summarized in Supplemental Table 3. Ki parametric maps from a single patient derived from Patlak a, basis function method irreversible (BFM-I) b, and BFM-I filtered for voxels showing a blood volume fraction above 50% c. The parametric map shows a coronal slice and a zoomed-in coronal slice of a tumour Fig. 3Scatter plots comparing Ki values from Patlak a and basis function method irreversible (BFM-I) b with those obtained by the weighted nonlinear regression (NLR) two-tissue irreversible model (2T3k) for the VOIs showing a preference for the irreversible two-tissue compartment model (2TC model). The different healthy organs and tumour tissue are indicated by different colours. The solid black line represents the line of identity (LoI), and the black dashed line represents the regression line Scatter plots comparing Ki values from Patlak a and basis function method irreversible (BFM-I) b with those obtained by the weighted nonlinear regression (NLR) two-tissue irreversible model (2T3k) for the VOIs showing a preference for the irreversible two-tissue compartment model (2TC model). The different healthy organs and tumour tissue are indicated by different colours. The solid black line represents the line of identity (LoI), and the black dashed line represents the regression line NLR 2T4k model preference with gPatlak & BFM-R Figure 4 illustrates the Ki parametric maps of a single patient, obtained using gPatlak (a), BFM-R (b) and BFM-R filtered for voxels showing a VB above 50% (c). Both parametric maps visually exhibited increased uncertainties and variability in parametric outcomes. The voxels showing a VB above 50% aligned with large blood pool structures and the BFM-R Ki parametric map became cleaner after removing these voxels gPatlak and BFM-R were compared again the reversible 2T4k model for all VOIs of the healthy organs and tumour lesions showing a preference for the 2T4k model. Figure 5a and b show the comparison between Ki values obtained from gPatlak and BFM-R with those derived by the 2T4k pharmacokinetic model. In one patient, the kidney could not be accurately fitted using the 2T4k model, likely due to motion-related artefacts, as the TAC showed evidence for motion-related problems by temporal irregularities in the TAC shape at the three last frames which cannot be explained by kinetics. The TAC of this patient is also illustrated in Supplemental Fig. 2. This data point was excluded from the regression fits and correlation coefficients. A strong correlation was found between Ki values obtained by gPatlak and the 2T4k model. However, gPatlak systematically underestimated Ki across all tissues, as also gPatlak does not account for VB. This also introduced a regression offset. BFM-R showed a strong correlation with the 2T4k pharmacokinetic model. The regression line was nearly parallel to the line of identity, with only a slight underestimation of Ki values, indicating close agreement. All correlation coefficients and regression parameters are summarized in Table 3. To examine the impact of applying an irreversible parametric model to tissue preferring reversible kinetics, we also compared Ki estimates from Patlak and BFM-I to those obtained by the 2T4k model. The correlation and regression fit coefficient of these analysis, based on healthy organs and tumour lesions, are illustrated in Supplemental Fig. 3 and summarized in Supplemental Table 2. Fig. 4Ki parametric maps from a single patient derived from gPatlak a, basis function method reversible (BFM-R) b, and BFM-R filtered for voxels showing a blood volume fraction above 50% c. The parametric map shows a coronal slice and a zoomed-in coronal slice of a tumour Ki parametric maps from a single patient derived from gPatlak a, basis function method reversible (BFM-R) b, and BFM-R filtered for voxels showing a blood volume fraction above 50% c. The parametric map shows a coronal slice and a zoomed-in coronal slice of a tumour Fig. 5Scatter plots comparing Ki values from gPatlak a and basis function method reversible (BFM-R) b with those obtained by the weighted nonlinear regression (NLR) two-tissue reversible model (2T4k). Panels a and b include a zoomed-out insert. The zoomed-in figures exclude one VOI with a very poor and unreliable model fit for better visualization. The different healthy organs and tumour tissue are indicated by different colours. The solid black line represents the line of identity (LoI), and the black dashed line represents the regression line Scatter plots comparing Ki values from gPatlak a and basis function method reversible (BFM-R) b with those obtained by the weighted nonlinear regression (NLR) two-tissue reversible model (2T4k). Panels a and b include a zoomed-out insert. The zoomed-in figures exclude one VOI with a very poor and unreliable model fit for better visualization. The different healthy organs and tumour tissue are indicated by different colours. The solid black line represents the line of identity (LoI), and the black dashed line represents the regression line K1 comparison between NLR & parametric models The parametric models BFM-I and BFM-R provide, in addition to Ki, also K1 estimates, as they incorporate VB as a kinetic parameter. Figure 6 illustrates the correlation of K1 values between BMF-I and the 2T3k model (a), and between BFM-R with the 2T4k model (b). Data points remain grouped according to their AIC-based model preference. A strong correlation was observed between BFM-I and the 2T3k model, with increased variability in the kidney and myocardium. Additionally, the spleen showed a slight underestimation in K1 estimates derived by BFM-I. BFM-R also showed a strong correlation with the NLR 2T4k model for K1. The regression line was nearly parallel to the line of identity, with a slight overestimation of K1 values. Notably, the kidney showed reduced variability and no evidence of underestimation in the spleen was found between BFM-R and the 2T4k model. All correlation coefficients and regression parameters are summarized in Table 3 Fig. 6Scatter plots comparing K1 values from basis function method irreversible (BFM-I) with the weighted nonlinear regression (NLR) two-tissue irreversible model (2T3k) a, and between with basis function method reversible (BFM-R) with the NLR two-tissue reversible model (2T4k). The different healthy organs and tumour tissue are indicated by different colours. The solid black line represents the line of identity (LoI), and the black dashed line represents the regression line Scatter plots comparing K1 values from basis function method irreversible (BFM-I) with the weighted nonlinear regression (NLR) two-tissue irreversible model (2T3k) a, and between with basis function method reversible (BFM-R) with the NLR two-tissue reversible model (2T4k). The different healthy organs and tumour tissue are indicated by different colours. The solid black line represents the line of identity (LoI), and the black dashed line represents the regression line Table 3Correlation and regression coefficients between the Ki and K1 values obtained through Patlak, gPatlak, BFM-I, and BFM-R with the Ki and K1 values obtained through weighed non-linear regression 2T3k and 2T4k pharmacokinetic modelsModel comparison R p-valueSlopeIntercept K i Patlak – NLR 2T3k0.97< 0.00010.85-0.00019BFM-I – NLR 2T3k0.99< 0.00011.00-0.00025gPatlak – NLR 2T4k0.90< 0.00010.84-0.0044BFM-R – NLR 2T4k0.95< 0.00010.99-0.0014 K 1 BFM-I – NLR 2T3k0.96< 0.00010.860.051BFM-R – NLR 2T4k0.96< 0.00010.980.045BFM-I, basis faction method irreversible; BFM-R, basis faction method reversible; NLR, non-linear regression; 2T3k, two-tissue irreversible model; 2T4k, two-tissue reversible model. Table 3Correlation and regression coefficients between the Ki and K1 values obtained through Patlak, gPatlak, BFM-I, and BFM-R with the Ki and K1 values obtained through weighed non-linear regression 2T3k and 2T4k pharmacokinetic modelsModel comparison R p-valueSlopeIntercept K i Patlak – NLR 2T3k0.97< 0.00010.85-0.00019BFM-I – NLR 2T3k0.99< 0.00011.00-0.00025gPatlak – NLR 2T4k0.90< 0.00010.84-0.0044BFM-R – NLR 2T4k0.95< 0.00010.99-0.0014 K 1 BFM-I – NLR 2T3k0.96< 0.00010.860.051BFM-R – NLR 2T4k0.96< 0.00010.980.045BFM-I, basis faction method irreversible; BFM-R, basis faction method reversible; NLR, non-linear regression; 2T3k, two-tissue irreversible model; 2T4k, two-tissue reversible model. Correlation and regression coefficients between the Ki and K1 values obtained through Patlak, gPatlak, BFM-I, and BFM-R with the Ki and K1 values obtained through weighed non-linear regression 2T3k and 2T4k pharmacokinetic models BFM-I, basis faction method irreversible; BFM-R, basis faction method reversible; NLR, non-linear regression; 2T3k, two-tissue irreversible model; 2T4k, two-tissue reversible model. Discussion This study aimed to evaluate the performance of various parametric models for voxel-wise kinetic analysis of dynamic 18F-FDG PET data acquired on a LAFOV PET/CT system. The performance of the parametric models was determined relative to 2TC models, which is considered the reference approach for describing the kinetics in dynamic 18F-FDG PET scans. However, complex models such as a 2TC model are not ideal for voxel-wise analyses as they are prone to convergence problems in the presence of noise [3]. Model preference for irreversible or reversible kinetics was first determined using AIC, and our results demonstrated that the liver, myocardium, lungs, and skeletal muscle preferred the irreversible 2T3k model. In contrast, model preference for the kidney, brain, spleen, and tumour lesions varied across patients, with some exhibiting irreversible and others reversible kinetics. These findings align with previous studies that have reported reversible 18F-FDG uptake in several organs [5–9]. However, whereas those studies suggested reversible kinetics predominantly in the liver, brain, lungs, and some tumour lesions, our data indicated reversible tracer kinetic behaviour primarily in the kidney, brain, spleen, and tumour lesions. Notably, we observed irreversible kinetics in the brain in only half of the patients, while Pantel et al. [9] reported decreased 18F-FDG uptake after 18.6 h in all patients. This suggests that a 70 min acquisition may not be sufficient to capture reversible kinetics in all patients. While not entirely overlapping, both findings underscore the need to evaluate the kinetic profile of these organs to select the appropriate parametric model. Simply applying Patlak linearization under the assumption of irreversible kinetics may lead to biased estimates of kinetic parameters, as was also shown before [6]. After establishing the correct 2TC model, we compared Ki values derived from Patlak and BFM-I with the 2T3k model for VOIs preferring irreversible kinetics. gPatlak and BFM-R were compared against the 2T4k model for VOIs having a preference for the reversible model. Patlak and gPatlak analyses showed strong correlations with the 2TC models but systematically underestimated Ki values due to their inability to correct for VB [3, 6, 11, 20]. This results in a bias for Patlak and gPatlak analyses, particularly for tissues with a larger VB [3, 6, 11, 20]. BFM-I and BFM-R also showed strong agreement with the 2TC models, with regression slopes of 1.0 and 0.99, respectively. BFM-R only showed a slight overestimation in Ki. Besides Ki, BFM-I and BFM-R also provided estimates of K1 [12]. K1 values from BFM-I correlated strongly with those obtained from the 2T3k model, although higher variability was observed in the kidney, myocardium, and certain tumour lesions. The spleen showed a slight underestimation in K1. K1 values derived from BFM-R also showed a strong relation with K1 derived by the 2T4k model, with less variability in the kidney and no apparent underestimation in the spleen, suggesting that BFM-R may be more accurate for highly perfused organs, as previous research demonstrated that modelling K1 values is harder in highly perfused organs [21]. However, it is complicated to directly compare the performance of BFM-I against that of BFM-R, as the data was already divided based on model preference. Overall, we found an excellent agreement between the BFM models and the 2TC models. This was expected given their mathematical equivalence. The main advantage of BFM over 2TC model fits is its use of predefined basis functions to enhance computation efficiency and robustness, particularly in parametric imaging [12]. Notably, agreement with 2TC model fits was generally better for tissues exhibiting irreversible kinetics. This was evident both in the correlation metrics and the parametric maps, where gPatlak and BFM-R displayed greater uncertainty compared to Patlak and BFM-I. The increased uncertainties for gPatlak and BFM-R are likely due to the inclusion of an additional parameter accounting for dephosphorylation, which inherently introduces more variance [6]. Nevertheless, Supplemental Figs. 1 and 2 and Supplemental Table 3 demonstrate that applying an inappropriate parametric model (e.g., using an irreversible model for tissues that prefer reversible kinetics, or vice versa) worsens parameter agreement, highlighting that mismatched models introduce more bias to the kinetic parameters. Several limitations in the current study should be acknowledged. First, we did not incorporate dual-input functions for the liver and lung, potentially influencing observed model preferences, as we found other preferences for the liver and lung compared to earlier studies [5–8]. Second, parametric imaging across multiple kinetic profiles was addressed through model selection based on the AIC at the organ level. However, model selection between BFM-I and BFM-R could alternatively be performed at the voxel level using voxel-wise AIC values, as previously demonstrated by Li et al. [22]. This approach may offer greater flexibility and could be explored in future studies, as it would eliminate the need for prior organ-based model selection. Third, largely uniform modeling settings were applied across all organs in this study. Previous work using 15O-water PET has shown that organ-specific kinetic models can yield improved performance [23]. Future research should therefore investigate the use of organ-specific parameter bounds to further optimize kinetic modeling for LAFOV PET/CT systems. In addition, complex model fitting is prone to convergence problems at voxel-by-voxel noise levels [3]. Although noise will likely remain a challenge in dynamic LAFOV PET imaging, previous research has shown that advanced reconstruction methods, including kernel-based methods, can reduce the impact of noise on complex model fitting [24]. Further research should therefore evaluate the potential benefits of additional noise strategies, such as a kernel reconstruction method [24]. Motion correction was not applied in this study. Prior research by Artesani et al. [25] showed the impact of patient motion on parametric imaging and Moradi et al. [2] showed that motion correction improved the accuracy of compartment modelling in LAFOV PET/CT imaging. Applying motion correction would have been particularly relevant in the three data points (one tumour, one kidney, and myocardium) where the 2TC model failed, likely due to movement. Finally, only 18F-FDG was evaluated. Nevertheless, the findings are expected to be transferable to other tracers, as Patlak and basis-function methods (gPatlak, BFM-I, and BFM-R) are broadly applicable. To transfer the data to other tracers, Patlak analysis requires the definition of the appropriate t* after which linearity of the method occurs and the basis-function methos requires the appropriate BF settings. Nonetheless, our findings demonstrate the importance of selecting the appropriate parametric model for dynamic PET, especially given the enhanced sensitivity and signal-to-noise ratios of LAFOV PET/CT systems. This may expose previously undetectable micro parameter contributions [26, 27]. Based on our results, we suggest first determining the kinetic profile of organs, such as the brain, kidney, spleen, and tumour lesions, that might show reversible kinetics. For organs showing irreversible kinetics, we recommend BFM-I as a parametric model due to its excellent agreement with the 2T3k model. Where reversible kinetics are present, BFM-R is preferred, due to its strong alignment with the 2T4k model. However, BFM-I and BFM-R require a full dynamic scan, limiting their use in clinical routine. In such cases, Patlak or gPatlak analyses may be more feasible in clinical routine, especially combined with ultra-short acquisitions and population-based input functions, as demonstrated by Sluis et al. [28] and Sari et al. [29]. Nevertheless, careful model selection remains essential and should depend on the required accuracy and precision in relation to the expected effect sizes. When accurate and precise quantification of the metabolic response needs to be captured as accurately as possible to capture small changes, Patlak and gPatlak are not recommended, and BFM-I or BFM-R should be used. However, the effect sizes are sufficiently large, shorter scan durations combined with Patlak or gPatlak analyses may be considered. Conclusions Our findings demonstrate the importance of selecting the appropriate parametric model for different healthy organs and tumour lesions to ensure accurate kinetic parameter estimation in dynamic 18F-FDG PET imaging using LAFOV PET/CT systems. While Patlak and gPatlak showed strong correlations with the 2TC models, they consistently underestimated Ki due to their inability to account for VB. These simplified models may remain useful in clinical settings where shorter acquisition times are required. However, when full dynamic scans are available, BFM-I and BFM-R are preferred due to their better agreement with the 2TC models and their ability to estimate both Ki and K1. As the availability of LAFOV PET/CT systems continues to grow, a tailored parametric imaging approach will be essential to enhance their potential in clinical and research settings. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1