 Original research
 Open access
 Published:
Simultaneous estimation of a modelderived input function for quantifying cerebral glucose metabolism with [^{18}F]FDG PET
EJNMMI Physics volumeÂ 11, ArticleÂ number:Â 11 (2024)
Abstract
Background
Quantification of the cerebral metabolic rate of glucose (CMRGlu) by dynamic [^{18}F]FDG PET requires invasive arterial sampling. Alternatives to using an arterial input function (AIF) include the simultaneous estimation (SIME) approach, which models the imagederived input function (IDIF) by a series of exponentials with coefficients obtained by fitting time activity curves (TACs) from multiple volumesofinterest. A limitation of SIME is the assumption that the input function can be modelled accurately by a series of exponentials. Alternatively, we propose a SIME approach based on the twotissue compartment model to extract a high signaltonoise ratio (SNR) modelderived input function (MDIF) from the wholebrain TAC. The purpose of this study is to present the MDIF approach and its implementation in the analysis of animal and human data.
Methods
Simulations were performed to assess the accuracy of the MDIF approach. Animal experiments were conducted to compare derived MDIFs to measured AIFs (nâ€‰=â€‰5). Using dynamic [^{18}F]FDG PET data from neurologically healthy volunteers (nâ€‰=â€‰18), the MDIF method was compared to the original SIMEIDIF. Lastly, the feasibility of extracting parametric images was investigated by implementing a variational Bayesian parameter estimation approach.
Results
Simulations demonstrated that the MDIF can be accurately extracted from a wholebrain TAC. Good agreement between MDIFs and measured AIFs was found in the animal experiments. Similarly, the MDIFtoIDIF areaunderthecurve ratio from the human data was 1.02â€‰Â±â€‰0.08, resulting in good agreement in grey matter CMRGlu: 24.5â€‰Â±â€‰3.6 and 23.9â€‰Â±â€‰3.2Â mL/100Â g/min for MDIF and IDIF, respectively. The MDIF method proved superior in characterizing the first pass of [^{18}F]FDG. Groupwise parametric images obtained with the MDIF showed the expected spatial patterns.
Conclusions
A modeldriven SIME method was proposed to derive high SNR input functions. Its potential was demonstrated by the good agreement between MDIFs and AIFs in animal experiments. In addition, CMRGlu estimates obtained in the human study agreed to literature values. The MDIF approach requires fewer fitting parameters than the original SIME method and has the advantage that it can model the shape of any input function. In turn, the high SNR of the MDIFs has the potential to facilitate the extraction of voxelwise parameters when combined with robust parameter estimation methods such as the variational Bayesian approach.
Background
Positron emission tomography (PET) imaging with 2deoxy2[^{18}F]fluoroDglucose ([^{18}F]FDG) has proven valuable for assessing cerebral energy metabolism in neurological diseases [1, 2]. Quantification of dynamic [^{18}F]FDG PET data is typically performed by measuring the cerebral metabolic rate of glucose (CMRGlu) by means of Patlak graphical analysis [3, 4], which requires measuring the timevarying concentration of [^{18}F]FDG in plasma (i.e., the arterial input function or AIF). Noninvasive alternatives to arterial sampling have primarily focused on extracting an imagederived input function (IDIF). PETonly methods of extracting the IDIF from the carotid arteries require careful correction of partial volume effects (PVE) [5, 6], which is typically performed by measuring the pointspread function of the PET system at the measurement location [7]. PVE corrections can be facilitated by combining magnetic resonance imaging (MRI) angiography of the feeding arteries [8], but PET and MRI misalignments can introduce errors in the IDIFâ€”which require sophisticated registration methods [9]â€”and complex vessel segmentation approaches can limit applicability [10]. Alternatively, hybrid PET/MR imaging allows for simultaneous acquisition of both functional and anatomical information, reducing misalignment errors [11, 12] and facilitating PVE correction [13, 14]. Although these approaches are promising, their complexity hinders the widespread use of IDIFs when evaluating [^{18}F]FDG PET images [15]. Lastly, acquiring dynamic [^{18}F]FDG PET data with a large axial fieldofview scanner may facilitate the localization of large arteries needed to extract the IDIF [16]; however, the limit access to such scanners fosters the development of simpler and readily available techniques to be implemented in any PET centre.
An alternative to extracting the IDIF from feeding arteries is the simultaneous estimation (SIME) approach. By assuming the input function is the same for all brain regions, the SIME method models the IDIF as a series of exponentials by which parameters are obtained by simultaneously fitting timeactivity curves (TACs) from various volumesofinterest (VOIs) and using blood samples as scalers [17,18,19]. Although promising, modelling the input function as a series of exponentials may not always be accurate [20], especially for the initial period following injection [21]. Strategies to overcome this limitation include maintaining a uniform experimental design across subjects and imaging sites (e.g., fixed injection duration), better characterizing the peak by incorporating the injection duration into the model [22], modelling the prepeak phase as a straight line [23], and using the IDIF first pass extracted with an MRbased method as a prior [24, 25].
This study investigated an alternative SIME approach in which the input function is defined by the irreversible twotissue compartment model (2TCM) used to characterize dynamic [^{18}F]FDG PET data, rather than by a series of exponentials. The modelderived input function (MDIF) is derived from the wholebrain (WB) TAC, given it has the highest signaltonoise ratio (SNR) obtainable in the PET brain data, with rate constants obtained by SIME. This approach requires fewer fitting parameters to define the input function compared to the original SIME approach, although late timepoint blood samples are still used as anchors as recommended when extracting IDIFs [6, 15]. The MDIF is free of the typical PVE that affect the IDIFs extracted from the feeding arteries and achieves low noise levels, while still retaining the subject and studyspecific shape of the input function, making the approach readily available and suitable for any experimental design and injection protocol. The overall objective of this work was to investigate the performance of the MDIF SIME approach.
Materials and methods
Modelderived input function
PET imaging of cerebral glucose metabolism with [^{18}F]FDG is typically based on the irreversible twotissue compartment model [3], in which [^{18}F]FDG enters the tissue via glucose transporters defined by an influx rate constant \({K}_{1}\) (in mL/g/min). Once in the first compartment (i.e., the free pool), [^{18}F]FDG can either return to the blood pool at an efflux rate defined by \({k}_{2}\) (in min^{âˆ’1}) or be phosphorylated at a rate defined by \({k}_{3}\) (in min^{âˆ’1}) and subsequently trapped in the metabolic pool. These processes are defined by the following two differential equations:
where \({C}_{1}\left(t\right)\) represents the activity concentration in the free pool and \({C}_{2}\left(t\right)\) in the metabolic pool. Eq. (3) provides the solution to Eqs. (1) and (2) for the total measured activity (i.e., \({C}_{PET}\left(t\right)=\left(1{V}_{b}\right)\left({C}_{1}\left(t\right)+{C}_{2}\left(t\right)\right)+{V}_{b}{C}_{b}\left(t\right)\)), which includes an additional term \({V}_{b}\) (in mL/g) to account for bloodborne [^{18}F]FDG activity [26]:
where \(*\) represents the convolution operation; \({C}_{p}\left(t\right)\) and \({C}_{b}\left(t\right)\) represent the plasma and wholeblood activity concentration of [^{18}F]FDG, respectively; and \({k}_{f}={k}_{2}+{k}_{3}\).
The previous equations can be combined and rearranged to derive an expression for the input function, i.e., the MDIF (see Appendix for full derivation). By selecting the WB TAC (\({C}_{wb}\left(t\right)\)) to define the MDIF, considering it has the highest SNR obtainable in a dynamic PET brain image, and assuming \({C}_{b}\left(t\right)=R{C}_{p}\left(t\right)\) (where \(R\) is the bloodtoplasma ratio), the MDIF is given by:
where \({\alpha }_{\mathrm{1,2}}=a\mp \sqrt{{a}^{2}b}\), with \(a=\frac{1}{2}\left(\frac{1{V}_{b}}{R{V}_{b}}{K}_{1}+{k}_{f}\right)\) and \(b=\frac{1{V}_{b}}{R{V}_{b}}{K}_{1}{k}_{3}\).
By employing this approach, the input function is defined by the four model parameters (\({K}_{1}, {k}_{2},{k}_{3}\) and \({V}_{b}\)) that characterize the WB TAC. Eq. (4) was developed for [^{18}F]FDG given it obeys the irreversible 2TCM and follows two key assumptions. First, [^{18}F]FDG does not require metabolite correction as the tracer and its metabolites are trapped in the cell and clearance of [^{18}F]FDG6phosphate is slow for acquisition times of up to 60Â min [27]. Second, the bloodtoplasma ratio was assumed to be a constant value for simplification purposes [28], although this assumption might not be valid for nonprimates [29]. The appendix provides the general solution for reversible radiotracers. We are currently developing a variation of the MDIF method for tracers that require metabolite correction, but this is beyond the scope of the current work.
MDIF SIME implementation
The SIME procedure [17] requires TACs from multiple volumesofinterest (VOIs) to derive the wholebrain kinetic parameters that define the MDIF (Eq. (4)). These TACs can be from anatomical VOIs or grouped based on functional similarity, as previously suggested by Wong et al., to identify TACs with distinct kinetics [18]. We implemented a kmeans clustering algorithm (explained below) to group TACs by functional similarity. After initial investigation, we observed that 3 clusters of each tissue type (i.e., grey [GM] and white matter [WM]) were sufficient to derive the MDIF, with littletono improvement observed by using more than 6 TACs (results not shown). A diagram outlining the MDIF SIME implementation is shown in Additional file 1: Fig. S1.
The parameters needed to compute the MDIF using Eq.Â (4) were obtained by simultaneously fitting six TACs to the irreversible 2TCM solution (Eq. (3)) using a nonlinear least squares (NLLS) fitting routine. The algorithm runs until the cost function, comprising of the residual sum of squares (RSS; in Eq. (5)), is minimized and the WB model parameters used to define the MDIF is obtained.
where \({C}_{j}\left(t\right)\) is the [^{18}F]FDGPET activity concentration of the j^{th} VOI, \({C}_{j}^{fit}\left(t\right)\) is the corresponding model estimate, \(n\) is the number of VOIs (not including the WB TAC), and \(T\) is the number of timeframes. Given the need for scalers (or anchors) to accurately obtain an input function with SIME [15], our procedure included two late timepoint blood samples to act as anchors. \({C}_{k}\) is the k^{th} blood sample activity concentration, and \({C}_{k}^{MDIF}\) is given by Eq. (4) at the same time as the blood sample.
The fitting procedure was performed in MATLAB (The MathWorks Inc., R2023a) using the optimization routine lsqnonlin and four parameters per TAC were included in the fitting routine (\({K}_{1}\), \({k}_{2}\), \({k}_{3}\), and \({V}_{b}\), for a total of 28 parameters). The upper bounds were set to 0.2Â mL/g/min, 0.4Â min^{âˆ’1}, 0.2Â min^{âˆ’1}, and 0.10Â mL/g for \({K}_{1}\), \({k}_{2}\), \({k}_{3}\), and \({V}_{b}\), respectively (based on literature values, see Additional file 1: Table S1). All lower bounds were set to 0.01. A constraint was added to the nonlinear optimization function implemented in this study to ensure the distribution volume (=â€‰\({K}_{1}/\left({k}_{2}+{k}_{3}\right)\)) is less than unity (i.e., \({K}_{1}<{k}_{2}+{k}_{3}\)), which was based on literature values (Additional file 1: Table S1).
kmeans clustering
For the animal and human studies, VOIs were generated by clustering the TACs based on functional similarity by implementing a kmeans clustering algorithm (Euclidean distance, 500Â iterations, 10Â replicates) that used the kmedoids MATLAB function kmediods [30]. For clustering only, PET images were denoised by applying the 3D highly constrained backprojection (HYPR3D) method [31], in which a Gaussian filter kernel of standard deviation 3Â voxels was used (equivalent to 6.3â€‰Ã—â€‰6.3â€‰Ã—â€‰6.1Â mm^{3} for the human PET data used in this study). Although HYPR3D introduces bias to the dynamic signal, which is known to affect quantification [32], this bias is not expected to influence the kmeans clustering algorithm. Instead, denoising with HYPR3D improves the clustering of similar TACs. Additional kmeans clustering implementation details are given in the appropriate sections below.
IDIF SIME implementation
For comparison to our MDIF SIME method, the original SIME approach based on the work by Feng et al. and Wong et al. [17, 18] was implemented to extract an IDIF by estimating its 7 parameters (Eq. (6)), while simultaneously fitting six VOI TACs (WB TAC not included) using a NLLS fitting routine. The same cost function from the MDIF SIME was used (Eq. (5)), except for an additional weight (=â€‰10) included in the blood samples portion of Eq. (5). All fitting was performed in MATLAB using the optimization routine lsqnonlin and four parameters per TAC were included in the fitting routine (\({K}_{1}\), \({k}_{2}\), \({k}_{3}\), and \({V}_{b}\); same constraint and bounds as described above) in addition to the 7 parameters required for the IDIF (Eq. (6)): \({A}_{i}\), \({\lambda }_{i}\), with iâ€‰=â€‰1,Â 2,Â 3, and a delay term \(\delta\) (total of 31 parameters). Upper bounds were 4000Â kBq/mL/min, 100Â kBq/mL, 50Â kBq/mL, 25Â min^{âˆ’1}, 1Â min^{âˆ’1}, and 0.1Â min^{âˆ’1}, for \({A}_{1}\), \({A}_{2}\), \({A}_{3}\), \({\lambda }_{1}\), \({\lambda }_{1}\), and \({\lambda }_{3}\), respectively. Lower bounds were set to zero. The delay term (\(\delta\)) was set to vary withinâ€‰Â±â€‰10Â s of an initial delay computed a priori based on the rise of the WB TAC.
Standalone fitting routine
To extract microparameters from anatomical VOIs, each TAC was fit to the irreversible 2TCM solution (Eq. (3)) using the MATLAB optimization routine lsqnonlin. Four parameters were included in the weighted NLLS (WNLLS) fitting routine (\({K}_{1}\), \({k}_{2}\), \({k}_{3}\), and \({V}_{b}\)). Upper bounds were set to 0.5Â mL/g/min, 0.5Â min^{âˆ’1}, 0.2Â min^{âˆ’1}, and 1Â mL/g for \({K}_{1}\), \({k}_{2}\), \({k}_{3}\), and \({V}_{b}\), respectively. All lower bounds were set to zero. The cost function was defined as a weighted RSS (WRSS), given by \(WRSS=\sum_{i=1}^{T}{w}_{i}{\left({C}_{j}^{fit}\left({t}_{i}\right){C}_{j}\left({t}_{i}\right)\right)}^{2}\). Weights for the fitting (\({w}_{i}\)) were defined as the inverse of variance of the PET measurement error (i.e., the ratio between the timeframe duration and the WB activity concentration, normalized to the maximum weight) [33]. A brain density of 1.05Â mL/g was used throughout the analyses.
Simulations
A theoretical AIF (\({C}_{p}\left(t\right)\)) was generated by Eq. (6) using \({A}_{1}\)Â =Â 850 (in arbitrary units [a.u.]/min), \({A}_{2}\)Â =Â 22 and \({A}_{3}\)Â =Â 21 (in a.u.), \({\lambda }_{1}\)Â =Â 4, \({\lambda }_{2}\)Â =Â 0.12 and \({\lambda }_{3}\)Â =Â 0.01 (in min^{âˆ’1}), and \(\delta\)Â =Â 0. These constants were obtained from Feng et al. and are based on experimental AIF measurements from a human study [34]. The theoretical AIF was used in the irreversible 2TCM solution (Eq. (3)) to generate simulated regional TACs for six theoretical VOIs (TableÂ 1). The microparameters shown in TableÂ 1 were chosen from a variety of studies [9, 35,36,37] and reflect the need for TACs with distinct kinetics for the SIME approach [18, 38]. The net clearance rate constant of [^{18}F]FDG (\({K}_{i}\)Â =Â \({K}_{1}{k}_{3}/{k}_{f}\), in mL/g/min), is also shown in TableÂ 1. A theoretical WB TAC was obtained by using the average microparameters from the six VOIs. Fig.Â 1 shows the theoretical AIF alongside simulated TACs. These curves and microparameters were considered the groundtruth for the simulations. Lastly, simulations used anchors extracted from the theoretical AIF at 28.5 and 53.5Â min timepoints, as indicated by the arrows in Fig.Â 1A.
To confirm the accuracy of the MDIF SIME routine, errors due to estimating the WB microparameters (Eq. (4)) with the SIME routine were evaluated. The MDIF was generated as described above and used to fit the six VOI TACs (TableÂ 1) to the irreversible 2TCM solution (Eq.Â (3)) in the standalone fitting routine. Then, the bestfit estimates of the four parameters (\({K}_{1}\), \({k}_{2}\), \({k}_{3}\), and \({V}_{b}\)) were compared to their input values.
Animal study
The MDIFs were assessed in comparison to AIFs using retrospective data from animal experiments [12] collected at the Lawson Health Research Institute. All experiments were conducted according to the regulations of the Canadian Council on Animal Care and approved by the Animal Care Committee at Western University. Data from juvenile Duroc pigs were collected on a 3Â T hybrid PET/MR scanner (Biograph mMR, Siemens Heathineers, Erlangen, Bavaria, Germany) using a 12channel PETcompatible receiver head coil. The five animals included in this study (weight, 21â€‰Â±â€‰2Â kg; injected [^{18}F]FDG activity, 90â€‰Â±â€‰20Â MBq; blood glucose level, 5.0â€‰Â±â€‰1.6Â mmol/L) were scanned under an anaesthetic combination of 1â€“3% isoflurane and 6â€“25Â mL/kg/h intravenous infusion of propofol. [^{18}F]FDG was administered via a cephalic vein immediately followed by a 60min acquisition of PET data in listmode. Arterial blood was continuously withdrawn from a femoral artery using an automated MRcompatible system (Swisstrace GmbH, Menzingen, Canton of Zug, Switzerland) at a sampling rate of 4Â mL/min for the first 5Â min, 1Â mL/min for the next 5Â min, and 0.5Â mL/min for the remaining 50Â min. Arterial blood samples were processed and converted into AIFs via rebinning to match the reconstructed PET time frames using pSample (PMOD Technologies LLC). Each measured AIF was corrected for dispersion and delay [39].
Dynamic PET images were reconstructed offline with the Siemens e7 tools into 51 timeframes (15â€‰Ã—â€‰2Â s, 6â€‰Ã—â€‰5Â s, 8â€‰Ã—â€‰15Â s, 4â€‰Ã—â€‰30Â s, 5â€‰Ã—â€‰60Â s, 5â€‰Ã—â€‰120Â s, 8â€‰Ã—â€‰300Â s) using a 3dimensional ordinary Poisson ordered subset expectation maximization (3DOPOSEM) method (3 iterations, 21 subsets) with corrections for decay, random coincidences, deadtime, detector normalization, data rebinning, attenuation, and scatter. Images were reconstructed into a 344â€‰Ã—â€‰344â€‰Ã—â€‰127 matrix with voxel size of 0.8â€‰Ã—â€‰0.8â€‰Ã—â€‰2.0Â mm and smoothed with a 3D gaussian filter of 4Â mm. Simultaneously to the PET acquisition, T_{1}weighted MR images were acquired (magnetizationprepared rapid gradientecho sequence [MPRAGE]; repetition/echo times [TR/TE], 2000/2.98Â ms; inversion time, 900Â ms; fieldofview [FoV], 256â€‰Ã—â€‰256Â mm^{2}; isotropic voxel size, 1Â mm^{3}; flip angle [Î±], 9Â°; 176 slices). At the end of the experiment, the animals were euthanized according to the animal care guidelines and transported to a computed tomography (CT) scanner to obtain a postmortem CTbased attenuation correction map.
For the kmeans clustering algorithm, a semiautomatic procedure was used to define a VOI encompassing the brain in each anatomical slice. The final WB VOI, which was the composite of all the slices, was used to generate seven clusters with the kmeans clustering approach as described above. The cluster containing mostly nonbrain regions (i.e., blood vessels and cerebrospinal fluid voxels) was excluded from the SIME algorithm. The SIME approach incorporated arterial samples collected at 20â€“30Â min and 40â€“60Â min postinjection as anchors. Although a constant bloodtoplasma ratio may not be valid for the porcine model used in this study [29], we assumed \(R\)Â =Â 1 throughout the analysis of the animal data.
Human study
Retrospective data from 18 neurologically healthy volunteers (44â€‰Â±â€‰15Â years, 77â€‰Â±â€‰17Â kg, 9Â M/9F; average injected activity, 180â€‰Â±â€‰40Â MBq [range 130â€“260Â MBq]; average injected activity per body weight, 2.43â€‰Â±â€‰0.39Â MBq/kg [range 1.80â€“3.70Â MBq/kg]; blood glucose level, 5.0â€‰Â±â€‰0.7Â mmol/L [range 4.1â€“6.8Â mmol/L]) collected at the Lawson Health Research Institute were used to assess the feasibility of obtaining MDIFs from dynamic [^{18}F]FDG PET data. The study was approved by the Western University Health Sciences Research Ethics Board and was conducted in accordance with the Declaration of Helsinki ethical standards. Participants provided written informed consent in compliance with the TriCouncil Policy Statement of Ethical Conduct for Research Involving Humans.
Scanning was performed on a 3Â T hybrid PET/MR scanner (Biograph mMR) using a 16channel PETcompatible coil (12 and 4channel head and neck coils, respectively). Each participant had their head immobilized during the 60min listmode PET acquisition performed following the [^{18}F]FDG bolus injection, which was immediately followed by saline flush. PET data were reconstructed offline with the Siemens e7 tools into 50 timeframes (15â€‰Ã—â€‰2Â s, 6â€‰Ã—â€‰5Â s, 8â€‰Ã—â€‰15Â s, 4â€‰Ã—â€‰30Â s, 5â€‰Ã—â€‰60Â s, 3â€‰Ã—â€‰120Â s, 8â€‰Ã—â€‰300Â s, and 1â€‰Ã—â€‰240Â s) using the iterative reconstruction algorithm 3DOPOSEM (3 iterations, 21 subsets, 3D gaussian filter of 2Â mm^{3}, and a zoom factor of 2). Dynamic PET images were reconstructed into a 172â€‰Ã—â€‰172â€‰Ã—â€‰127 matrix with voxel size of 2.1â€‰Ã—â€‰2.1â€‰Ã—â€‰2.0Â mm (FoV, 359â€‰Ã—â€‰359â€‰Ã—â€‰258Â mm^{3}). Corrections for decay, random coincidences, deadtime, detector normalization, data rebinning, attenuation, and scatter were performed. MRbased attenuation correction was performed with a vendorprovided ultrashort echo time MRI sequence (TR, 11.94Â ms; TE, 0.07 and 2.46Â ms; Î±, 10Âº; FoV, 300â€‰Ã—â€‰300â€‰Ã—â€‰300Â mm^{3}; voxel size, 1.6â€‰Ã—â€‰1.6â€‰Ã—â€‰1.6Â mm^{3}). Motion correction was applied to the time frames 24â€“50 (corresponding to 1.5 to 60Â min) by realigning each frame to the mean image using statistical parametric mapping (v12, SPM12; https://www.fil.ion.ucl.ac.uk/spm/software/spm12/). Each session also included acquiring an MPRAGE image (TR/TE, 2400/2.25Â ms; Î±, 8Â°; FoV, 205â€‰Ã—â€‰205Â mm^{2}; 240 slices; voxel size, 0.8â€‰Ã—â€‰0.8â€‰Ã—â€‰0.8Â mm^{3}), which was coregistered to the PET space using SPM12.
The MPRAGE image was segmented into six tissue classes using SPM12. A WB mask was created by combining GM and WM tissue probability maps (80% threshold; cerebrospinal fluid voxels excluded). Five clusters for each tissue class (i.e., GM and WM) were generated by implementing the kmeans clustering algorithm as described above. Out of the 5 clusters, the 3 largest ones were chosen for the SIME algorithms. Two venous samples collected during dynamic PET imaging (one 20â€“30Â min postinjection and the other 45â€“60Â min postinjection) were used as scalers when arterialtovenous [^{18}F]FDG equilibrium was assumed [6, 40]. The bloodtoplasma ratio was assumed to beâ€‰~â€‰0.9 (\(R\)Â =Â 0.9) to derive the MDIF from the human data [28].
Variance of estimated \({{\varvec{K}}}_{{\varvec{i}}}\)
Coefficient of variation (CV) of \({K}_{i}\) estimates obtained by fitting each cluster TAC to the 2TCM solution was computed as \(CV\left(\%\right)=100\sqrt{{\sigma }^{2}}/{K}_{i}\), where the variance (\({\sigma }^{2}\)) was obtained from the error propagation rule: \({\sigma }^{2}=\left[\begin{array}{ccc}\frac{\partial {K}_{i}}{\partial {K}_{1}}& \frac{\partial {K}_{i}}{\partial {k}_{2}}& \frac{\partial {K}_{i}}{\partial {k}_{3}}\end{array}\right]\gamma \left({K}_{1},{k}_{2},{k}_{3}\right){\left[\begin{array}{ccc}\frac{\partial {K}_{i}}{\partial {K}_{1}}& \frac{\partial {K}_{i}}{\partial {k}_{2}}& \frac{\partial {K}_{i}}{\partial {k}_{3}}\end{array}\right]}^{T}\), where the Hessian (covariance) matrix (\(\gamma\)) was approximated as \(\gamma ={\left({J}^{T}J\right)}^{1}WRSS/T\) (\(J\) is the Jacobian matrix from the MATLAB function lsqnonlin and \(T\) the number of time frames).
Goodness of fit
Following the derivation of the MDIF and IDIF with their respective SIME algorithms, goodness of fit of each input function was assessed via the WRSS from the standalone fitting routine. For this, each input function was used to fit the GM and WM TACs, and WRSS was calculated for the first 3Â min of data, as well as for the entire 60Â min of data. In addition, fitting to the irreversible 2TCM solution was evaluated qualitatively for the GM and WM TACs of one representative subject, as well as for TACs from two randomly selected GM and WM voxels.
Regional measurements
Anatomical MPRAGE images were normalized to the Montreal Neurological Institute (MNI; McGill University, Montreal, Quebec, Canada) space using SPM12 and the same deformation field was applied to the dynamic PET data. Average TACs were extracted for GM and WM (99% threshold; subject space), as well as twelve VOIs (frontal, occipital, parietal, and temporal lobes, as well as insula, cingulate, hippocampus, precuneus, caudate nucleus, putamen, thalamus, and cerebellum; MNI space) using the automated anatomical labelling atlas (Wake Forest University PickAtlas, http://fmri.wfubmc.edu/cms/software). The mask generated for the twelve VOIs was multiplied by the GM mask (80% threshold; MNI space) to extract the central portion of the VOIs, avoiding PVE at the boundaries.
Patlak analysis (\(t>{t}^{*}\)Â =Â 20Â min) was performed for each VOI, from which the slope (\({K}_{i}\); in mL/100Â g/min) was converted to CMRGlu (in Âµmol/100Â g/min) by assuming a lumped constant of 0.52 [35] and incorporating measurements of the blood concentration of glucose (in Âµmol/mL).
Voxelwise parametric images of microparameters were generated by incorporating the MDIF into the variational Bayesian approach proposed by Castellaro et al. [33]. Specifically for this step, the kmeans clustering algorithm described above was used to generate a functional atlas containing 40 clusters for each subject: 20 GM, 10 WM and 10 nonbrain clusters, the latter including cerebrospinal fluid and voxels surrounding the brain. In the variational Bayesian fitting routine, each cluster TAC was fit to the irreversible 2TCM solution, with its rate constants used as priors for all intracluster voxels assuming a Gaussian distribution of possible values. Parametric maps of CMRGlu were computed from \({K}_{i}\) images (=â€‰\({K}_{1}{k}_{3}/\left({k}_{2}+{k}_{3}\right)\)) [35]. Resulting parametric images (i.e., CMRGlu, \({K}_{1}\), \({k}_{2}\), \({k}_{3}\), and \({V}_{b}\)) were normalized to the MNI space using SPM12.
Statistics
Area under the curve (AUC) was used to assess the similarity between input functions. Percent error was calculated as \(100\left(m\widehat{m}\right)/\widehat{m}\), where \(m\) and \(\widehat{m}\) are the observed and expected values, respectively. Percent difference between two observed measurements \({m}_{1}\) and \({m}_{2}\) was computed as \(100\left{m}_{1}{m}_{2}\right/\overline{m }\), where \(\overline{m }\) is the average measurement. A repeated measures twoway ANOVA was used to evaluate if estimates were affected by input function (MDIF vs. IDIF) and VOI; multiple comparisons tests were used to evaluate differences between MDIF and IDIF estimates. Linear regression was used to compare MDIF and IDIF results, from which the lineofbest fit was obtained alongside the 95% confidence intervals (CIs) for both the slope and intercept. Correlation was assessed by means of the Pearson correlation coefficient (\(\rho\)). All datasets were found to be normally distributed with normal QQ plots. Statistical tests were performed using GraphPad Prism (version 9, GraphPad Software, San Diego, California USA, www.graphpad.com). Statistical significance was defined by \(\alpha\)Â =Â 0.05. Measurements are expressed in terms of meanâ€‰Â±â€‰one standard deviation alongside the CIs in square brackets when relevant.
Results
Simulations
FigureÂ 2A shows the MDIF obtained with the SIME algorithm and Fig.Â 2B the difference between the simulated (true) input function and the MDIF. Average error of microparameters were 0.33â€‰Â±â€‰0.01% for \({K}_{1}\), 0.31â€‰Â±â€‰0.07% for \({k}_{2}\),â€‰âˆ’â€‰0.13â€‰Â±â€‰0.04% for \({k}_{3}\), 0.32â€‰Â±â€‰0.01% for \({V}_{b}\), and 0.05â€‰Â±â€‰0.01% for \({K}_{i}\). These values were computed from the 6 VOIs used in the MDIF SIME approach. Regional \({K}_{i}\) estimates obtained with the MDIF (3.26â€‰Â±â€‰0.89Â mL/100Â g/min) were nearly identical to their input values.
Animal study
Average WB estimates of \({K}_{1}\), \({k}_{2}\), \({k}_{3}\), and \({V}_{b}\) obtained using the AIF were 0.098â€‰Â±â€‰0.021Â mL/g/min, 0.148â€‰Â±â€‰0.034Â min^{âˆ’1}, 0.013â€‰Â±â€‰0.009Â min^{âˆ’1}, and 0.060â€‰Â±â€‰0.019Â mL/g, respectively (nâ€‰=â€‰5). Corresponding average WB \({K}_{i}\) was 0.86â€‰Â±â€‰0.65Â mL/100Â g/min. Good MDIFtoAIF agreement was observed across the five animals with an AUC ratio of 0.99â€‰Â±â€‰0.06 (difference of 5.1â€‰Â±â€‰2.2%; AUC_{05}â€‰=â€‰0.93â€‰Â±â€‰0.14, AUC_{510}â€‰=â€‰0.92â€‰Â±â€‰0.09, AUC_{1030}â€‰=â€‰1.05â€‰Â±â€‰0.08, AUC_{3060}â€‰=â€‰0.99â€‰Â±â€‰0.07, where the subscript refers to the time period in minutes). Percent difference between the arterial blood measurements used as anchors and the MDIF was 6.8â€‰Â±â€‰4.1% for the 20â€“30min anchor and 8.9â€‰Â±â€‰6.6% for the 40â€“60min anchor. Average WB estimates of the fitting parameters obtained using the MDIF were \({K}_{1}\)Â =Â 0.123â€‰Â±â€‰0.028Â mL/g/min (pâ€‰=â€‰0.048), \({k}_{2}\)Â =Â 0.195â€‰Â±â€‰0.060Â min^{âˆ’1} (pâ€‰=â€‰0.058), \({k}_{3}\)Â =Â 0.013â€‰Â±â€‰0.004Â min^{âˆ’1} (pâ€‰=â€‰0.958), and \({V}_{b}\)Â =Â 0.056â€‰Â±â€‰0.016Â mL/g (pâ€‰=â€‰0.691). No significant difference was observed in average WB \({K}_{i}\) estimates obtained with the MDIF (0.85â€‰Â±â€‰0.48Â mL/100Â g/min, pâ€‰=â€‰0.911) compared to AIF measurements. FigureÂ 3 shows a comparison between the average measured AIF and MDIF curves.
Human study
MDIFs were successfully extracted from all 18 subjects (average plasma curve is shown in Fig.Â 4A). Good overall agreement was observed between plasma MDIFs and IDIFs, with an AUC ratio of 1.02â€‰Â±â€‰0.08 (difference of 6.3â€‰Â±â€‰4.6%; AUC_{05}â€‰=â€‰1.06â€‰Â±â€‰0.24, AUC_{510}â€‰=â€‰1.06â€‰Â±â€‰0.20, AUC_{1030}â€‰=â€‰1.04â€‰Â±â€‰0.10, AUC_{3060}â€‰=â€‰0.97â€‰Â±â€‰0.05). Percent difference between the venous samples and the MDIFs was 4.2â€‰Â±â€‰3.3% for the 20â€“30min anchor (29.3â€‰Â±â€‰2.6Â min) and 9.3â€‰Â±â€‰9.0% for the 45â€“60min anchor (55.2â€‰Â±â€‰1.7Â min). Likewise, differences of 1.6â€‰Â±â€‰2.0% and 2.4â€‰Â±â€‰2.5% were found between the venous samples and the SIME IDIFs values at the two time points. FigureÂ 4B presents the average plasma IDIF obtained with the original SIME approach (i.e., Eq.Â (6)). FigureÂ 4C shows the plasma MDIF and IDIF (series of exponentials) from one representative subject, alongside the two venous samples used as anchors for all three input functions. In addition, Fig.Â 4D shows the corresponding WB [^{18}F]FDGTAC used to define the MDIF, Fig.Â 4E the cluster TACs used in both SIME procedures, and Fig.Â 4F the anatomical localization of the clusters presented in Fig.Â 4E. Average IDIF parameters obtained from the SIME routine were 2941â€‰Â±â€‰1045Â kBq/min, 6.6â€‰Â±â€‰6.8Â kBq, 9.2â€‰Â±â€‰2.5Â kBq, 10.1â€‰Â±â€‰1.9Â min^{âˆ’1}, 0.32â€‰Â±â€‰0.34Â min^{âˆ’1}, 0.012â€‰Â±â€‰0.004Â min^{âˆ’1}, and 26â€‰Â±â€‰7Â s, for \({A}_{1}\), \({A}_{2}\), \({A}_{3}\), \({\lambda }_{1}\), \({\lambda }_{2}\), \({\lambda }_{3}\), and \(\delta\), respectively.
The microparameters obtained from the two SIME approaches are summarized in Additional file 1: Table S2, alongside the CV values for each cluster. Average WB MDIF parameters were \({K}_{1}\)Â =Â 0.103â€‰Â±â€‰0.019Â mL/g/min, \({k}_{2}\)Â =Â 0.153â€‰Â±â€‰0.055Â min^{âˆ’1}, \({k}_{3}\)Â =Â 0.050â€‰Â±â€‰0.029Â min^{âˆ’1}, and \({V}_{b}\)Â =Â 0.067â€‰Â±â€‰0.012Â mL/g, with a CV of 0.4â€‰Â±â€‰0.1%. For comparison, respective IDIF estimates from the standalone fitting routine were \({K}_{1}\)Â =Â 0.109â€‰Â±â€‰0.024Â mL/g/min (pâ€‰=â€‰0.473), \({k}_{2}\)Â =Â 0.145â€‰Â±â€‰0.048Â min^{âˆ’1} (pâ€‰=â€‰0.633), \({k}_{3}\)Â =Â 0.040â€‰Â±â€‰0.012Â min^{âˆ’1} (pâ€‰=â€‰0.095), and \({V}_{b}\)Â =Â 0.043â€‰Â±â€‰0.010Â mL/g (pâ€‰<â€‰0.001), with a CV of 4.9â€‰Â±â€‰2.0%. Compared to MDIF SIME resulting clusters parameters, IDIF SIME results had higher estimates of \({K}_{1}\) (difference of approx. 9%) and \({k}_{2}\) (~â€‰4%), and lower estimates of \({k}_{3}\) (~â€‰16%) and \({V}_{b}\) (~â€‰30%), although significance was only observed for CBV (pâ€‰<â€‰0.001 for all clusters).
Goodness of fit
WRSS, indicating goodness of fit, for the first 3Â min of each GM TAC was two orders of magnitude lower (pâ€‰<â€‰0.001) for the TACs fit using the MDIFs compared to corresponding TACs fit using the IDIFs (Fig.Â 5A). Likewise, a significantly lower WRSS for curves fitted with the MDIFs was also observed for the entirety of data (pâ€‰<â€‰0.001; Fig.Â 5B). Average WRSS for the first 3Â min of data was (2.1â€‰Â±â€‰1.2)â€‰Ã—â€‰10^{âˆ’3} and (1.1â€‰Â±â€‰0.9)â€‰Ã—â€‰10^{âˆ’1}Â (kBq/mL)^{2} for the GM TACs fitted with the MDIFs and IDIFs, respectively (nâ€‰=â€‰18); corresponding average WRSS for the 60Â min of data was (5.6â€‰Â±â€‰3.3)â€‰Ã—â€‰10^{âˆ’3} and (2.7â€‰Â±â€‰2.3)â€‰Ã—â€‰10^{âˆ’1}Â (kBq/mL)^{2}. The stronger goodness of fit obtained with the MDIF is also demonstrated in Fig.Â 5CF, which shows the fit of GM and WM TACs to the irreversible 2TCM solution. Results are for data from the same representative subject shown in Fig.Â 4.
Regional measurements
No significant differences between the two methods (i.e., MDIF and IDIF) were observed for either \({K}_{i}\) or CMRGlu estimates for anatomical VOIs, except for the cerebellum (pâ€‰=â€‰0.030 and pâ€‰=â€‰0.034 for \({K}_{i}\) and CMRGlu, respectively). Macro and microparameter estimates for anatomical VOIs obtained with the MDIF are summarized in TableÂ 2. For comparison, total GM and WM \({K}_{i}\) estimates obtained using the IDIFs were 2.51â€‰Â±â€‰0.50 (pâ€‰=â€‰0.008) and 1.12â€‰Â±â€‰0.22Â mL/100Â g/min (pâ€‰=â€‰0.139), respectively. Respective CMRGlu estimates were 23.9â€‰Â±â€‰3.2 (pâ€‰=â€‰0.010) and 10.7â€‰Â±â€‰1.3Â Âµmol/100Â g/min (pâ€‰=â€‰0.142). FigureÂ 6 shows the linear regression comparing IDIF and MDIFderived CMRGlu estimates. Average (Â±â€‰standard error) regression for GM had a slope of 0.85â€‰Â±â€‰0.06 [0.71â€“0.98] and an intercept of 3.1â€‰Â±â€‰1.6Â Âµmol/100Â g/min [âˆ’â€‰0.2 to 6.5Â Âµmol/100Â g/min] (R^{2}â€‰=â€‰0.92, \(\rho\)Â =Â 0.96, pâ€‰<â€‰0.001). Similarly, average (Â±â€‰standard error) regression for WM had a slope of 0.82â€‰Â±â€‰0.06 [0.69 to 0.95] and an intercept of 1.7â€‰Â±â€‰0.7Â Âµmol/100Â g/min [0.2 to 3.1Â Âµmol/100Â g/min] (R^{2}â€‰=â€‰0.92, \(\rho\)Â =Â 0.96, pâ€‰<â€‰0.001). Lastly, Fig.Â 7 shows average CMRGlu and microparameter images obtained with the variational Bayesian approach using MDIFs (nâ€‰=â€‰18); respective images obtained using the IDIFs are shown in Additional file 1: Fig. S2; differences between microparameters are shown in Additional file 1: Fig. S3.
Discussion
This study investigated a SIME approach to generate an input function that replaced the commonly used series of exponentials by a modelderived equation based on the irreversible twotissue compartment model. The simulations demonstrated that the input function can be recovered accurately by the SIME routine, with errors in macro and microparameters of less than 1%. The accuracy of the MDIF approach predicted by the simulations was confirmed in the animal study in which the MDIFtoAIF AUC ratio was close to unity and the input functions demonstrated good agreement (Fig.Â 3). The discrepancy between MDIF and AIF peaks could be partially explained by the assumption of a constant bloodtoplasma ratio of one, which might not be accurate in this animal model. For instance, Somogyi observed that the entirety of true sugars in the blood of pigs was present in plasma [41], although we could not find any studies that measured the bloodtoplasma ratio of [^{18}F]FDG in this species. Future studies are needed to measure the [^{18}F]FDG bloodtoplasma for this animal model.
Good agreement between MDIFs and IDIFs was observed when the two SIME approaches were applied to dynamic [^{18}F]FDG PET data from healthy human participants (Fig.Â 4). No significant differences between \({K}_{i}\) or CMRGlu estimates were found, most likely since both methods used the same anchors. Since the MDIF is derived directly from the tissue TAC, and not restricted to gamma distribution functions as the original SIME, it is applicable to any injection protocol. The flexibility of the method is evident in Fig.Â 5. The injection protocol in this study involved a flush of saline that resulted in an inadvertent second peak during the early phase of the tissue TACs. This unexpected shape was only properly characterized by the MDIF method as evident by the improved fit of the tissue TACs to the 2TCM solution (Fig.Â 5Câ€“F). A further advantage of the MDIF approach is the number of fitting parameters required to characterize the input function was reduced from seven (i.e., Eq. (6)) to four. The higher number of parameters needed to characterize the IDIF leads to multiple local minima during the minimization procedure, making the IDIF SIME procedure more sensitive to the chosen starting values. Consequently, higher variance of the estimated macroparameter (\({K}_{i}\)) was observed when deriving the IDIF with the SIME routine (CV ofâ€‰~â€‰5%) compared to MDIF SIME (~â€‰2%).
As with all SIME methods, the MDIF approach requires TACs with varying kinetics and anchors to stabilize the minimization routine. The former was achieved using the kmeans clustering algorithm, as previously suggested [42]. For the latter, either one or two blood samples were sufficient to properly capture the shape of the MDIF; however, the presence of noise in experimental PET data could introduce errors in the procedure. Care should be taken when selecting the appropriate number of anchors since previous studies observed that the use of a single scaling point was not sufficient to calibrate IDIFs [9, 18]. A potential limitation with anchors is that measurement errors will propagate through the SIME procedure. Increasing the number of venous samples could reduce these errors. In addition, consideration should be given when selecting the timewindow to collect venous blood, as arteriovenous equilibrium occurs at different times for different tracers [43]. Although a completely noninvasive approach would be ideal, as proposed previously [23, 44], arterial or venous samples should always be considered to improve quantification accuracy.
The MDIF method has several advantages over other methods proposed to extract an IDIF. Compared to vessel segmentation approaches, it is free of PVEs and can produce input functions with greater SNR given the MDIF is derived from the WB TAC, while segmentation approaches often suffer from lower SNR due to the limited vessel volume used to extract the IDIF. Furthermore, IDIFs extracted from carotid arteries tend to overestimate their tails and scaling them with blood samples is recommended for quantification [15]. Additionally, the MDIF method can be used for any PET scanner given that the segmentation portion (i.e., kmeans clustering; Fig.Â 4F) can be implemented by using a WB mask obtained from PETonly data. Thus, there is no need for specific anatomical data, such as timeofflight MR imagesâ€”often required in standard IDIF extraction methods. A further advantage of the MDIF approach is that it accounts for subjectspecific variations in metabolism, which is a known limitation of populationbased input functions. These advantages place the MDIF SIME approach at a unique position to be implemented in any PET centre. However, it may be necessary to incorporate a distributed parameter model to better character the vascular phase of the tracer considering recent advances in PET imaging are enabling the temporal resolution to be reduced to 1 or 2Â s [45].
Although no arterial sampling was available to validate the MDIF SIME approach in humans, the CMRGlu values and rate constants obtained in the human study were in agreement with literature estimates [9, 35,36,37, 46,47,48] (Additional file 1: Table S1). Future studies involving human participants aimed at validating the MDIF SIME method would also provide the opportunity to optimize the technique by investigating the use of other parameter estimation methodsâ€”such as the simulated annealing method [49] implemented by Ogden et al. to minimize the SIME cost function [38]. As it is, the derived MDIF better characterized the first pass of [^{18}F]FDG when compared to the IDIF modelled as a series of exponentials (Fig.Â 5).
The most obvious application of the MDIF would be to estimate CMRGlu from dynamic [^{18}F]FDG PET data; however, there is growing interest in understanding how glucose metabolism is potentially affected by diseaserelated alterations in glucose delivery and phosphorylation [50, 51]. Altered microparameters have been observed in gliomas [52], suggesting they could improve differential diagnosis of malignant brain tumors [53]. Likewise, evidence of cerebrovascular contributions to Alzheimerâ€™s disease has led to the hypothesis that hypometabolism characteristic of neurodegeneration is related, in part, to impaired glucose transport across the bloodâ€“brain barrier (i.e., \({K}_{1}\)) [51]. These examples highlight the value of optimizing key steps in quantifying glucose metabolism to better understand major neurological diseases. However, accurately estimating the model rate constants at a voxelwise level is challenging [54]. Quantification of microparameters by kinetic modelling suffers from substantial statistical uncertainties due to the high noise level affecting voxel TACs. Efforts to develop image denoising approaches to improve quantification accuracy include postreconstruction methods, such as the 4dimensional iterative highly constrained backprojection algorithm [55], and during image reconstruction, such as physicsinformed artificial intelligence reconstruction algorithms trained to increase the SNR [56]. Additionally, the accuracy of microparameters is linked to the choice of parameter estimation method [57, 58]. Lastly, the measured AIF, required for proper quantification, is an inherently noisy procedure that is often fit with a mathematical model to generate a continuous noiseless version for kinetic analysis [22]. The ability to generate highquality input functions with the modeldriven SIME approach could improve the reliability of microparameter estimates (Fig.Â 5), especially when combined with the variational Bayesian [33] parameter estimation method. In the parametric images obtained in this study (Fig.Â 7), \({K}_{1}\) maps presented the expected pattern of higher delivery of [^{18}F]FDG to GM tissue.
Although a more sophisticated implementation is required, the MDIF SIME approach described here can be extended to other radiotracers that follow a reversible 2TCM (see Appendix), including when there are concerns regarding the irreversibility of [^{18}F]FDG. Sari et al. implemented the IDIF SIME approach to evaluate dynamic data acquired with a serotonin receptor tracer and observed similar results in comparison to those obtained with the measured AIF [25]. Bartlett et al. and Ogden et al. used venous samples as scalers in the IDIF SIME method for a variety of radiotracers, including [^{18}F]FDG, and reported good agreement to results obtained when arterial blood scalers were used, albeit with slightly biased estimates [38, 43]. Like with other SIME methods, only the plasma input function is recovered, while correcting for vascular contributions requires the wholeblood input function [38]. Further work is required to investigate if the MDIF method can be extended to simultaneously estimate the wholeblood input function while performing metabolite correction. Finally, similar to methods proposed for measuring cerebral blood flow [59, 60], an MDIF equation for radiotracers that follow the onetissue compartment model could be derived following the steps outlined in the Appendix.
Conclusion
In this study, we presented a SIME method based on the irreversible 2TCM to derive a high SNR input function. The extracted MDIF was compared to AIF curves in animal experiments and good MDIFtoAIF agreement was observed. In the human study, MDIF macroparameters (i.e., \({K}_{i}\) and CMRGlu) estimates were in agreement with estimates obtained using the IDIF SIME method, which models the input function by a series of exponentials. Rate constants obtained with the MDIF were in good agreement with literature values. The MDIF approach has the advantages that it can model the shape of any input function and requires fewer fitting parameters. Additionally, combining the derived MDIF with the variational Bayesian approach allowed for the generation of parametric images of kinetic parameters.
Availability of data and materials
Scripts used in this study are available for download in an open repository (a link will be provided upon acceptance). The datasets generated and/or analysed during the current study are not publicly available due to data privacy but are available from the corresponding author on reasonable request.
Abbreviations
 [^{18}F]FDG:

2Deoxy2[^{18}F]fluoroDglucose
 2TCM:

Twotissue compartment model
 3D:

3Dimensional
 3DOPOSEM:

3D ordinary Poisson ordered subset expectation maximization
 AIF:

Arterial input function
 a.u.:

Arbitrary units
 AUC:

Area under the curve
 CI:

Confidence interval
 CMRGlu:

Cerebral metabolic rate of glucose
 CT:

Computed tomography
 CV:

Coefficient of variation
 FoV:

Fieldofview
 GM:

Grey matter
 HYPR3D:

3D highly constrained backprojection
 IDIF:

Imagederived input function
 MDIF:

Modelderived input function
 MNI:

Montreal Neurological Institute
 MPRAGE:

Magnetizationprepared rapid gradientecho
 MRI:

Magnetic resonance imaging
 NLLS:

Nonlinear least squares
 PET:

Positron emission tomography
 PVE:

Partial volume effects
 RSS:

Residual sum of squares
 SIME:

Simultaneous estimation
 SUV:

Standardized uptake value
 TAC:

Timeactivity curve
 TE:

Echo time
 TR:

Repetition time
 WB:

Whole brain
 WM:

White matter
 WNLLS:

Weighted nonlinear least squares
 WRSS:

Weighted residual sum of squares
 VOI:

Volumeofinterest
References
Johnson KA, Fox NC, Sperling RA, Klunk WE. Brain imaging in Alzheimer disease. Cold Spring Harb Perspect Med. 2012;2:a006213â€“a006213.
SchÃ¶ll M, DamiÃ¡n A, Engler H. Fluorodeoxyglucose PET in neurology and psychiatry. PET Clin. 2014;9:371â€“90.
Berti V, Vanzi E, Polito C, Pupi A. Back to the future: the absolute quantification of cerebral metabolic rate of glucose. Clin Transl Imaging. 2013;1:289â€“96.
Patlak CS, Blasberg RG, Fenstermacher JD. Graphical evaluation of bloodtobrain transfer constants from multipletime uptake data. J Cereb Blood Flow Metab. 1983;3:1â€“7.
Croteau E, LavallÃ©e Ã‰, Labbe SM, Hubert L, Pifferi F, Rousseau JA, et al. Imagederived input function in dynamic human PET/CT: methodology and validation with 11Cacetate and 18F fluorothioheptadecanoic acid in muscle and 18Ffluorodeoxyglucose in brain. Eur J Nucl Med Mol Imaging. 2010;37:1539â€“50.
Chen K, Bandy D, Reiman E, Huang SC, Lawson M, Feng D, et al. Noninvasive quantification of the cerebral metabolic rate for glucose using positron emission tomography, 18Ffluoro2deoxyglucose, the Patlak method, and an imagederived input function. J Cereb Blood Flow Metab. 1998;18:716â€“23.
Asselin MC, Cunningham VJ, Amano S, Gunn RN, Nahmias C. Parametrically defined cerebral blood vessels as noninvasive blood input functions for brain PET studies. Phys Med Biol. 2004;49:1033â€“54.
Fung EK, Carson RE. Cerebral blood flow with [15O]water PET studies using an imagederived input function and MRdefined carotid centerlines. Phys Med Biol. 2013;58:1903â€“23.
Sari H, Erlandsson K, Law I, Larsson HBW, Ourselin S, Arridge S, et al. Estimation of an image derived input function with MRdefined carotid arteries in FDGPET human studies using a novel partial volume correction method. J Cereb Blood Flow Metab. 2017;37:1398â€“409.
ZanottiFregonara P, Maroy R, Comtat C, Jan S, Gaura V, BarHen A, et al. Comparison of 3 methods of automated internal carotid segmentation in human brain PET studies: application tot he estimation of arterial input function. J Nucl Med. 2009;50:461â€“7.
Anazodo U, Kewin M, Finger E, Thiessen J, Hadway J, Butler J, et al. Preliminary evaluation of MRIderived input function for quantitative measurement of glucose metabolism in an integrated PETMRI. EJNMMI Phys. 2015;2:A80.
Dassanayake P, Cui L, Finger E, Kewin M, Hadaway J, Soddu A, et al. caliPER: a software for bloodfree parametric Patlak mapping using PET/MRI input function. Neuroimage. 2022;256: 119261.
Khalighi MM, Deller TW, Fan AP, Gulaka PK, Shen B, Singh P, et al. Imagederived input function estimation on a TOFenabled PET/MR for cerebral blood flow mapping. J Cereb Blood Flow Metab. 2018;38:126â€“35.
Sundar LKS, Muzik O, Rischka L, Hahn A, Rausch I, Lanzenberger R, et al. Towards quantitative [18F]FDGPET/MRI of the brain: automated MRdriven calculation of an imagederived input function for the noninvasive determination of cerebral glucose metabolic rates. J Cereb Blood Flow Metab. 2019;39:1516â€“30.
ZanottiFregonara P, Chen K, Liow JSS, Fujita M, Innis RB. Imagederived input function for brain PET studies: many challenges and few opportunities. J Cerebr Blood Flow Metabol. 2011;31:1986â€“98.
Sari H, Eriksson L, Mingels C, Alberts I, Casey ME, AfsharOromieh A, et al. Feasibility of using abbreviated scan protocols with populationbased input functions for accurate kinetic modeling of [18F]FDG datasets from a long axial FOV PET scanner. Eur J Nucl Med Mol Imaging. 2023;50:257â€“65.
Feng D, Wong KP, ChiMing Wu, Siu WC. A technique for extracting physiological parameters and the required input function simultaneously from PET image measurements: theory and simulation study. IEEE Trans Inform Tech Biomed. 1997;1:243â€“54.
Wong KP, Feng D, Meikle SR, Fulham MJ. Simultaneous estimation of physiological parameters and the input function  In vivo PET data. IEEE Trans Inf Technol Biomed. 2001;5:67â€“76.
Feng DD, Chen K, Wen L. Noninvasive input function acquisition and simultaneous estimations with physiological parameters for PET quantification: a brief review. IEEE Trans Radiat Plasma Med Sci. 2020;4:676â€“83.
van der Weijden CWJ, Mossel P, Bartels AL, Dierckx RAJO, Luurtsema G, Lammertsma AA, et al. Noninvasive kinetic modelling approaches for quantitative analysis of brain PET studies. Eur J Nucl Med Mol Imaging. 2023;50:1636â€“50.
YiGen W. Noninvasive quantification of local cerebral metabolic rate of glucose for clinical application using positron emission tomography and 18 FFluoro2Deoxy d Glucose. J Cereb Blood Flow Metab. 2008;28:242â€“50.
Tonietto M, Rizzo G, Veronese M, Borgan F, Bloomfield PS, Howes O, et al. A unified framework for plasma data modeling in dynamic positron emission tomography studies. IEEE Trans Biomed Eng. 2019;66:1447â€“55.
Zanderigo F, Ogden RT, Parsey RV. Noninvasive bloodfree full quantification of positron emission tomography radioligand binding. J Cereb Blood Flow Metab. 2015;35:148â€“56.
Sari H, Erlandsson K, Barnes A, Atkinson D, Arridge S, Ourselin S, et al. Exploiting an MRI derived arterial input function to improve the PET simultaneous estimation method: Validation of assumptions. 2014 IEEE Nuclear Science Symposium and Medical Imaging Conference, NSS/MIC 2014. Institute of Electrical and Electronics Engineers Inc.; 2016.
Sari H, Erlandsson K, Marner L, Law I, Larsson HBW, Thielemans K, et al. Noninvasive kinetic modelling of PET tracers with radiometabolites using a constrained simultaneous estimation method: evaluation with 11CSB201745. EJNMMI Res. 2018;8.
Lammertsma AA. Compartmental Modeling in Emission Tomography. In: Handbook of Particle Detection and Imaging. Berlin and Heidelberg: Springer; 2012. p. 1065â€“81.
Phelps ME, Huang SC, Hoffman EJ, Selin C, Sokoloff L, Kuhl DE. Tomographic measurement of local cerebral glucose metabolic rate in humans with (F18)2fluoro2deoxyDglucose: Validation of method. Ann Neurol. 1979;6:371â€“88.
Nahmias C, Wahl LM, Amano S, Asselin MC, Chirakal R. Equilibration of 6[18F]fluoroLmtyrosine between plasma and erythrocytes. J Nucl Med. 2000;41:1636â€“41.
Buxton DB. Noninvasive measurement of mouse myocardial glucose uptake with 18 FFDG. J Nucl Med. 2014;55:866.1866.
Park HS, Jun CH. A simple and fast algorithm for Kmedoids clustering. Expert Syst Appl. 2009;36:3336â€“41.
Christian BT, Vandehey NT, Floberg JM, Mistretta CA. Dynamic PET denoising with HYPR processing. J Nucl Med. 2010;51:1147â€“54.
Floberg JM, Mistretta CA, Weichert JP, Hall LT, Holden JE, Christian BT. Improved kinetic analysis of dynamic PET data with optimized HYPRLR. Med Phys. 2012;39:3319â€“31.
Castellaro M, Rizzo G, Tonietto M, Veronese M, Turkheimer FE, Chappell MA, et al. A variational bayesian inference method for parametric imaging of PET data. Neuroimage. 2017;150:136â€“49.
Feng D, Huang SC, Wang X. Models for computer simulation studies of input functions for tracer kinetic modeling with positron emission tomography. Int J Biomed Comput. 1993;32:95â€“110.
Reivich M, Alavi A, Wolf A, Fowler J, Russell J, Arnett C, et al. Glucose metabolic rate kinetic model parameter determination in humans: The lumped constants and rate constants for [18F]fluorodeoxyglucose and [11C]deoxyglucose. J Cereb Blood Flow Metab. 1985;5:179â€“92.
ZanottiFregonara P, Fadaili EM, Maroy R, Comtat C, Souloumiac A, Jan S, et al. Comparison of eight methods for the estimation of the imagederived input function in dynamic [18F]FDG PET human brain studies. J Cereb Blood Flow Metab. 2009;29:1825â€“35.
Huisman MC, van Golen LW, Hoetjes NJ, Greuter HN, Schober P, Ijzerman RG, et al. Cerebral blood flow and glucose metabolism in healthy volunteers measured using a highresolution PET scanner. EJNMMI Res. 2012;2:1â€“9.
Ogden RT, Zanderigo F, Choy S, Mann JJ, Parsey RV. Simultaneous estimation of input functions: an empirical study. J Cereb Blood Flow Metab. 2010;30:816â€“26.
Meyer E. Simultaneous correction for tracer arrival delay and dispersion in CBF measurements by the H215O autoradiographic method and dynamic PET. J Nucl Med. 1989;30:1069â€“78.
Wakita K, Imahori Y, Ido T, Fujii R, Horii H, Shimizu M, et al. Simplification for measuring input function of FDG PET: investigation of 1point blood sampling method. J Nucl Med. 2000;41:1484â€“90.
Somogyi M. The distribution of sugar and rate of glycolysis in the blood of some mammals. J Biol Chem. 1933;103:665â€“70.
Wong KP, Feng D, Meikle SR, Fulham MJ. Noninvasive extraction of physiological parameters in quantitative PET studies using simultaneous estimation and cluster analysis. In: IEEE Nuclear science symposium and medical imaging conference. 2000.
Bartlett EA, Ananth M, Rossano S, Zhang M, Yang J, Lin S, fei, et al. Quantification of positron emission tomography data using simultaneous estimation of the input function: validation with venous blood and replication of clinical studies. Mol Imaging Biol. 2019;21:926â€“34.
Roccia E, Mikhno A, Ogden RT, Mann JJ, Laine AF, Angelini ED, et al. Quantifying brain [18F]FDG uptake noninvasively by combining medical health records and dynamic PET imaging data. IEEE J Biomed Health Inform. 2019;23:2576â€“82.
St Lawrence KS, Lee TY. An adiabatic approximation to the tissue homogeneity model for water exchange in the brain: I. theoretical derivation. J Cerebr Blood Flow Metabol. 1998;118:1365â€“77.
Huang SC, Phelps ME, Hoffman EJ, Sideris K, Selin CJ, Kuhl DE. Noninvasive determination of local cerebral metabolic rate of glucose in man. Am J Physiol. 1980;238(1):E62â€“82.
Sasaki H, Kanno I, Murakami M, Shishido F, Uemura K. Tomographic mapping of kinetic rate constants in the fluorodeoxyglucose model using dynamic positron emission tomography. J Cereb Blood Flow Metab. 1986;6:447â€“54.
Mosconi L, Tsui WH, Rusinek H, De Santi S, Li Y, Wang GJ, et al. Quantitation, regional vulnerability, and kinetic modeling of brain glucose metabolism in mild Alzheimerâ€™s disease. Eur J Nucl Med Mol Imaging. 2007;34:1467â€“79.
Kirkpatrick S, Gelatt CD, Vecchi MP. Optimization by simulated annealing. Science. 1979;1983(220):671â€“80.
Hermanides J, Hong YT, Trivedi M, Outtrim J, Aigbirhio F, Nestor PJ, et al. Metabolic derangements are associated with impaired glucose delivery following traumatic brain injury. Brain. 2021;144:3492â€“504.
Sweeney MD, Montagne A, Sagare AP, Nation DA, Schneider LS, Chui HC, et al. Vascular dysfunctionâ€”The disregarded partner of Alzheimerâ€™s disease. Alzheimerâ€™s Dementia. 2019;15:158â€“67.
Ishikawa M, Kikuchi H, Nishizawa S, Yonekura Y. Evaluation of glucose transport in malignant glioma by PET. Brain Edema VIII. Vienna: Springer; 1980. p. 165â€“7.
Kimura N, Yamamoto Y, Kameyama R, Hatakeyama T, Kawai N, Nishiyama Y. Diagnostic value of kinetic analysis using dynamic 18FFDGPET in patients with malignant primary brain tumor. Nucl Med Commun. 2009;30:602â€“9.
Hallett WA. Quantification in clinical fluorodeoxyglucose positron emission tomography. Nucl Med Commun. 2004;25:647â€“50.
Bevington CWJ, Cheng JC, Sossi V. A 4D iterative HYPR denoising operator improves PET image quality. IEEE Trans Radiat Plasma Med Sci. 2022;6:641â€“55.
Reader AJ, Corda G, Mehranian A, da CostaLuis C, Ellis S, Schnabel JA. Deep learning for PET image reconstruction. IEEE Trans Radiat Plasma Med Sci. 2020;5:1â€“25.
Ben BF, Vauchot F, MarianoGoulart D. Comparative assessment of linear leastsquares, nonlinear leastsquares, and Patlak graphical method for regional and local quantitative tracer kinetic modeling in cerebral dynamic 18FFDG PET. Med Phys. 2019;46:1260â€“71.
Ikoma Y, Watabe H, Shidahara M, Naganawa M, Kimura Y. PET kinetic analysis: error consideration of quantitative analysis in dynamic studies. Ann Nucl Med. 2008;22:1â€“11.
Narciso L, Ssali T, Liu L, Jesso S, Hicks JW, Anazodo U, et al. Noninvasive quantification of cerebral blood flow using hybrid PET/MR imaging to extract the [15O]H2O imagederived input function free of partial volume errors. J Magn Reson Imaging. 2022;56:1243â€“55.
Kudomi N, Maeda Y, Yamamoto Y, Nishiyama Y. Reconstruction of an input function from a dynamic PET water image using multiple tissue curves. Phys Med Biol. 2016;61:5755â€“67.
Acknowledgements
We thank Dr. Alessandra Bertoldo (University of Padova) for sharing the variational Bayesian MATLAB scripts to generate the parametric images.
Funding
This work was supported by the Canadian Institutes of Health Research [Grant No. 148600 and Project Grant No. 202109PJT] and the Alzheimerâ€™s Drug Discovery Foundation [Grant No. 7767312].
Author information
Authors and Affiliations
Contributions
LN contributed to the theory, simulations, data analysis, data interpretation, MATLAB scripts development, and writing of the manuscript. GD contributed to data analysis. PD contributed to data acquisition and analysis. LL contributed to data acquisition, PET image reconstruction, and was responsible for the arterial sampling system. SP contributed to data analysis. UA contributed to conception, design of the work, and data acquisition. AS contributed to conception, design of the work, and funding support. KStL contributed to conception, design of the work, interpretation of data, substantively revised the manuscript, and funding support. All authors reviewed and agreed to the manuscript content. All authors read and approved the final manuscript.
Corresponding author
Ethics declarations
Ethics approval and consent to participate
Retrospective data from animal experiments collected at the Lawson Health Research Institute were used in this study. Experiments were conducted according to the regulations of the Canadian Council on Animal Care and approved by the Animal Care Committee at Western University (approval file number 2018â€“052). Retrospective data from neurologically healthy volunteers collected at the Lawson Health Research Institute were used in the analysis. The original study was approved by the Western University Health Sciences Research Ethics Board (approval file number 105895) and was conducted in accordance with the Declaration of Helsinki ethical standards. Participants provided written informed consent in compliance with the TriCouncil Policy Statement of Ethical Conduct for Research Involving Humans.
Consent for publication
Participants provided written informed consent before entry in the study.
Competing interests
The authors declare that they have no competing interests.
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Supplementary Information
Additional file 1.
Supplementary Figures and Tables.
Appendix
Appendix
The derivation of the MDIF equation (Eq. (4)) begins by combining the differential equations from the irreversible 2TCM (Eqs. (1) and (2)) after applying the Laplace Transform. Thus, the activity concentration in the first (\(\overline{{C }_{1}}\left(s\right)\)), second (\(\overline{{C }_{2}}\left(s\right)\)), and total (\(\overline{{C }_{T}}\left(s\right)\)) compartments are given by:
where \(s\) is the Laplace variable. For this step, the following Laplace transform was used: \({L}\left\{{f}{\prime}\left(t\right)\right\}=s\overline{F }\left(s\right)f\left(0\right)\), given \({C}_{1}\left(0\right)={C}_{2}\left(0\right)=0\).
As PET measurements (\(\overline{{C }_{PET}}\left(s\right)\)) include both tissue and blood activity contributions (i.e., \({C}_{PET}\left(t\right)=\left(1{V}_{b}\right){C}_{T}\left(t\right)+{V}_{b}{C}_{b}\left(t\right)\)), an additional term must be included to account for blood volume (\({V}_{b}\)) activity concentration originated from the wholeblood input function (\({C}_{b}\left(t\right)=R{C}_{p}\left(t\right)\), where \(R\) is the bloodtoplasma ratio):
Including Eqs.Â (9) in (10) yields:
which can be rewritten as:
where \({\alpha }_{\mathrm{1,2}}=a\mp \sqrt{{a}^{2}b}\), with \(a=\frac{1}{2}\left(\frac{1{V}_{b}}{R{V}_{b}}{K}_{1}+{k}_{f}\right)\), and \(b=\frac{1{V}_{b}}{R{V}_{b}}{K}_{1}{k}_{3}\).
Next, the MDIF can be obtained by isolating \(\overline{{C }_{p}}\left(s\right)\) in Eq.Â (12), which requires partial fraction decomposition, and the resulting equation is given by:
The solution for \({C}_{p}^{MDIF}\left(t\right)\) given by Eq. (4) is obtained by applying the inverse Laplace Transform to Eq.Â (13), considering: \({{L}}^{1}\left\{\frac{1}{s+a}\right\}={e}^{at}\) and \({{L}}^{1}\left\{\overline{F }\left(s\right)\bullet \overline{G }\left(s\right)\right\}=f\left(t\right)*g\left(t\right)\), where \(*\) is the convolution operator.
The MDIF, \({C}_{p}^{MDIF}\left(t\right)\), for reversible radiotracers is obtained by solving for \(\overline{{C }_{p}}\left(s\right)\) following the steps outlined above:
where \({\alpha }_{\mathrm{1,2}}=a\mp \sqrt{{a}^{2}4b}\) \(a=\frac{1}{2}\left(\frac{1{V}_{b}}{R{V}_{b}}{K}_{1}+{k}_{f}+{k}_{4}\right)\), and \(b=\frac{1{V}_{b}}{R{V}_{b}}{K}_{1}\left({k}_{3}+{k}_{4}\right)+{k}_{2}{k}_{4}\). Note, here we considered radiotracers of constant bloodtoplasma ratio and no metabolite production. The inclusion of metabolite correction would require a more sophisticated solution, which is outside the scope of this work.
Rights and permissions
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://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Narciso, L., Deller, G., Dassanayake, P. et al. Simultaneous estimation of a modelderived input function for quantifying cerebral glucose metabolism with [^{18}F]FDG PET. EJNMMI Phys 11, 11 (2024). https://doi.org/10.1186/s40658024006146
Received:
Accepted:
Published:
DOI: https://doi.org/10.1186/s40658024006146