Single-tracer [¹⁸F]FDG PET quantification of blood–brain barrier permeability and cerebral blood flow: Validation using dual-tracer PET

Abstract

Imaging the blood–brain barrier (BBB) permeability of molecular PET tracers may allow pathway-specific assessment of the diverse transport mechanisms expressed at the BBB. However, PET quantification of BBB permeability typically requires dual-tracer protocols involving a flow tracer, increasing methodological complexity and clinical burden. We evaluated a single-tracer dynamic PET method for quantifying the BBB permeability-surface area (PS) of [¹⁸F]fluorodeoxyglucose (FDG) against the conventional dual-tracer method. Our method uses high-temporal resolution imaging (1 s/frame), an image-derived arterial input function, and distributed kinetic modelling to simultaneously estimate cerebral blood flow (CBF) and transvascular transport rate K₁, from which BBB PS is calculated via the Renkin-Crone equation. Single-tracer and conventional dual-tracer PS estimates were compared in 18 volunteers scanned with the flow-tracer [¹¹C]butanol and [¹⁸F]FDG PET. Single-tracer [¹⁸F]FDG PS estimates had 3.7% mean bias and 4.2% standard deviation of differences compared to dual-tracer estimates. This was enabled by strong agreement between [¹⁸F]FDG and [¹¹C]butanol CBF estimates (Pearson R = 0.85, p < 0.001; mean difference: −0.057 ± 0.078 mL/min/cm³). These results demonstrate that high-temporal resolution dynamic PET enables single-tracer quantification of both CBF and BBB PS without a dedicated flow tracer, expanding opportunities for quantitative studies of molecular BBB transport across a broad range of tracers and disorders.

Introduction

Blood–brain barrier (BBB) permeability imaging has contributed to our basic understanding of the BBB’s role in neurological ¹ and systemic disease pathology. ² In neuroimaging, the BBB is typically regarded as a structural barrier that selectively excludes molecules, and permeability imaging methods to probe non-specific BBB leakage are well established. ³ However, the BBB comprises a wide range of mechanisms that facilitate molecular transport from blood into the brain. ¹ The ability to measure molecular permeability through these BBB pathways may provide new understandings on the basic function and role of molecular transport in neurological and systemic disorders.

BBB permeability can be assessed in humans with dynamic tracer imaging and kinetic modelling, and the tracer used determines the mechanism of BBB transport that is assessed. Intravascular probes such as gadolinium-based contrast agents for dynamic contrast-enhanced (DCE) MRI,^4,5 iodinated contrast agents for CT perfusion, ⁶ or [⁶⁸Ga]EDTA in PET ⁷ can assess non-specific BBB leakage under the assumption that they do not effectively extravasate at an intact BBB. ³ Investigators have also used molecular PET tracers towards probing more specific BBB transport mechanisms, such as endothelial sodium-potassium pumps with the potassium analogue [⁸²Rb]Cl. ⁸ The large catalogue of molecular PET tracers offers the potential to more specifically assess the various transport systems and mechanisms expressed at the BBB.

However, it is challenging to measure the BBB permeability of PET tracers by conventional compartment modelling. Compartment modelling allows direct estimation of BBB transport rate K₁, which reflects a mixture of cerebral blood flow (CBF) and the tracer’s permeability-surface area (PS) product, a kinetic measure of BBB permeability. K₁ approximates PS for very low tracer extraction fractions (E; i.e. PS ≪ CBF) and has been used to measure BBB permeability for subtle leakages seen in ageing ⁵ as well as for other tracers with low BBB permeability.^7,8

More generally for tracers with higher extraction fractions, quantification of BBB PS requires separate measures of CBF and K₁. In addition to using K₁ for subtle BBB leakage, DCE-MRI ⁴ and CT perfusion ⁶ can estimate PS by obtaining both CBF and K₁ from a single dynamic scan due to their ability to perform high-temporal resolution imaging (1–3 s/frame) and distributed kinetic modelling^9,10 or model-free deconvolution.^4,11 This has not been possible with conventional PET systems due to their inadequate sensitivity and inability to perform high-temporal resolution imaging. As such, molecular BBB permeability studies required a challenging procedure involving two consecutive dynamic PET scans and a flow tracer (e.g. [¹⁵O]water, [¹¹C]butanol), limiting their broader utilization.

Recent advances in high-sensitivity PET systems have enabled high-temporal resolution dynamic imaging (≈1 s/frame) and measurement of CBF based on modelling the vascular transit of the PET tracer.^11–13 Using this principle, we recently developed a single-scan method to non-invasively measure the BBB permeability of various molecular PET tracers without the need for a flow tracer or arterial blood sampling. ¹² Our method uses total-body PET ¹⁴ for high-temporal resolution imaging and simultaneous imaging of the brain and of cardiac blood pools to non-invasively obtain a high-quality image-derived arterial input function (IDIF). While our BBB PS and CBF estimates fell within expected ranges, it is unclear whether single-tracer estimates quantitatively agree with the conventional dual-tracer PET method. Furthermore, our original method analyzed the first 2 min of dynamic data, and it is unclear whether these truncated time-activity curves affect the quantification of CBF, K₁ and PS. In this work, we investigated methodological factors affecting the quantification of BBB PS and CBF using [¹⁸F]fluorodeoxyglucose (FDG) as a representative example and compared single-tracer versus dual-tracer estimates derived with the flow tracer [¹¹C]butanol.

Material and methods

Data acquisition and image reconstruction

This study was approved by the institutional review board at the University of California, Davis (IRB #1921252; Clinicaltrials.gov identifier NCT06014515) and written informed consent was obtained for all human participants prior to study procedures. All study procedures were conducted following guidelines outlined in the Belmont Report. Eighteen participants (mean age: 53 ± 19 years; 11 (61%) females) received dynamic PET imaging on the 194 cm axial field-of-view uEXPLORER total-body PET/CT system (United Imaging Healthcare, Shanghai, China) ¹⁴ using both [¹¹C]butanol and [¹⁸F]FDG. Ten-minute dynamic [¹¹C]butanol PET (mean injected activity: 274.8 ± 8.8 MBq) was obtained first and followed by 60-min dynamic [¹⁸F]FDG PET (364.2 ± 16.9 MBq) after at least a 2-h interval to allow [¹¹C] decay ([¹¹C] half-life: 20.4 min). Fifteen participants underwent a same-day protocol (mean interval: 2.65 ± 0.13 h) and three with a 2-day protocol within 14 days (range: 4–11 days). Nine participants were healthy volunteers with no self-reported history of major disease within the past 5 years while the remaining nine had a history of cardiometabolic disease.

Image reconstruction was performed using a vendor implementation of a standard time-of-flight ordered-subset expectation-maximization algorithm with four iterations, 20 subsets and no point spread function modelling. Standard corrections for attenuation, scatter, randoms, dead time and radionuclide decay were applied. ¹⁴ High-temporal resolution reconstructions for the first 3 min (100 frames: 60 × 1, 30 × 2, 10 × 6 s) were generated for all dynamic scans. For [¹⁸F]FDG, we also generated 60-min dynamic reconstructions at standard temporal resolution (29 frames: 6 × 10, 2 × 30, 6 × 60, 5 × 120, 4 × 180, 6 × 300 s). We combined 3-min high-temporal resolution data with the last 57-min standard-temporal resolution data to create high-temporal resolution 60-min dynamic datasets (120 frames: 60 × 1, 30 × 2, 10 × 6, 5 × 60, 5 × 120, 4 × 180, 6 × 300 s). Two separate reconstructions were needed for 60-min high-temporal resolution dynamic [¹⁸F]FDG datasets because the vendor reconstruction software did not allow more than 100 frames/series. For [¹¹C]butanol, only the first 2 min of data were analyzed to minimize the contribution of radiometabolites. ¹⁵

Tracer kinetic modelling of PS

The Renkin–Crone equation defines an exponential relationship between the tracer extraction fraction (E) and the ratio of PS and CBF^16,17:

$$E=1-e^{-PS/CBF}.$$ (1)

Although the original Renkin–Crone model assumed negligible backflux, Larson et al. showed that equation (1) remains valid when backflux is modelled explicitly. ¹⁰ The BBB transport rate K₁ is then given by:

$$K_1=CBF\cdot E.$$ (2)

Equation (2) shows that K₁ approximates CBF when E is close to 1 (PS ≫ CBF) as with freely-diffusible flow tracers like [¹¹C]butanol or [¹⁵O]water. When E is small (PS ≪ CBF), $E\approx PS/CBF$ and $K_1\approx PS$ , an approximation used to estimate BBB permeability for slowly extracted tracers^7,8 and in subtle BBB leakage. ⁵ Rearranging equation (1) with equation (2) yields a general expression for PS:

$$PS=-CBF\cdot \ln (1-K_1/CBF).$$ (3)

Standard compartment modelling

The conventional approach for estimating the BBB PS of [¹⁸F]FDG uses tracer kinetic modelling of two dynamic PET scans: one with a flow tracer to measure CBF and a second with [¹⁸F]FDG to measure K₁. Dynamic PET time-activity curves are commonly modelled as a linear shift-invariant system:

$$Q\left(t\right)=C_a\left(t-t_d\right)\otimes R(t),$$ (4)

where $C_a(t)$ is the arterial input function, $t_d$ is the time delay between tracer arrival at the measured arterial input function and the local arterial supply, ⊗ is the convolution operator and $R(t)$ is the impulse response function (IRF) of the tracer kinetic model. Using total-body PET, a whole-blood IDIF can be obtained from the ascending aorta for brain studies. The IRF of the standard two-tissue compartment (S2TC) model commonly used for [¹⁸F]FDG is:

$$R^{S2TC}(t)=\{\begin{array}{cc}{v}_b,& t=0,\\ G{e}^{-\alpha t}+H{e}^{-\beta t},& t>0,\end{array}$$ (5)

where $G,$ $H,$ α, β are combinations of microkinetic parameters $K_1,$ k₂, k₃, and k₄ ¹⁸ :

$$\begin{array}{l}{K}_1=G+H,k_2=\frac{G\alpha +H\beta }{G+H},\\ k_3=\frac{GH{\left(\alpha -\beta \right)}^2}{\left(G+H\right)\left(G\alpha +H\beta \right)},\\ k_4=\frac{\left(G+H\right)\alpha \beta }{G\alpha +H\beta }.\end{array}$$

Evaluating equation (4) with equation (5) yields:

$$\begin{array}{l}{Q}^{S2TC}\left(t\right)=v_b{C}_{wb}\left(t-t_d\right)+\\ C_p\left(t-t_d\right)\otimes \\ \left(G{e}^{-\alpha t}+H{e}^{-\beta t}\right),\end{array}$$ (6)

where $C_{wb}(t)$ and $C_p(t)$ are the whole-blood and plasma arterial input functions, respectively. For [¹⁸F]FDG, the whole-blood IDIF was converted to plasma concentration by dividing by the whole-blood-to-plasma ratio $BPR\left(t\right)=0.97-0.06e^{-0.085*t}$ determined by Naganawa et al. ¹⁹

For a freely-diffusible flow tracer like [¹¹C]butanol, the standard one-tissue compartment model (S1TC) is commonly used. ¹⁵ The IRF of the S1TC model is:

$$R^{S1TC}(t)=\{\begin{array}{cc}{v}_b,& t=0,\\ K_1{e}^{-k_{2}t},& t>0.\end{array}$$ (7)

For [¹¹C]butanol, we used $C_p\left(t\right)=C_{wb}(t)$ as it rapidly and uniformly distributes between plasma and erythrocytes. ²⁰ While [¹¹C]butanol is slowly metabolized in the body, using only the first 2 min of the time-activity curve mitigates the effect of radiometabolites. ¹⁵

High-temporal resolution kinetic modelling

To avoid the need for a flow tracer, we previously developed a high-temporal resolution early kinetic modelling approach, utilizing a rapidly sampled IDIF and tissue time-activity curves, to simultaneously estimate CBF and K₁ from [¹⁸F]FDG and other tracers. ¹² Specifically, we used the adiabatic approximation to the tissue homogeneity (AATH) model, ⁹ which provides an analytical time-domain solution to a distributed kinetic model ¹⁰ described by a system of partial differential equations. The AATH IRF is:

$$R^{AATH}(t)=\{\begin{array}{cc}CBF,& 0\le t<T_c,\\ K_1{e}^{-k_2(t-T_c)},& t\ge T_c,\end{array}$$ (8)

where $T_c$ is the mean vascular transit time for the tracer to traverse the entire vascular volume within a voxel or region of interest. The AATH IRF comprises a fast vascular phase $(0\le t<T_c)$ describing the intravascular transit of the tracer followed by a tissue phase $(t\ge T_c)$ with a response similar to the S1TC IRF but shifted in time by $T_c.$ We have shown that the S1TC model is a special case of the AATH model when $T_c$ tends to 0, consistent with the interpretation that the S1TC model effectively assumes instantaneous vascular mixing.^12,13

However, because the AATH model effectively contains only one tissue compartment, it cannot adequately fit the full 60-min time-activity curve of targeted PET tracers like [¹⁸F]FDG. Previously, this limited our analysis to the first 2 min of dynamic data, ¹² during which the signal contribution of [¹⁸F]FDG phosphorylation is small. Whether this empirically defined analysis time window affects CBF, K₁ or PS estimates is unclear.

Here we therefore extended the AATH model to include two tissue compartments. By adding a tissue compartment, the vascular phase remains unchanged, but the tissue phase gains an additional exponential term, analogous to the S2TC IRF tissue terms but shifted in time by $T_c.$ The resulting flow-modified two-tissue compartment (F2TC) model IRF is:

$$R^{F2TC}(t)=\{\begin{array}{cc}CBF,& 0\le t<T_c,\\ G{e}^{-\alpha (t-T_c)}+H{e}^{-\beta (t-T_c)},& t\ge T_c,\end{array}$$ (9)

where G, H, α and β are related to K₁, k₂, k₃ and k₄ in the same manner as in the S2TC model. Evaluating equation (4) with equation (9) yields:

$$\begin{array}{l}{Q}^{F2TC}\left(t\right)=CBF\left[\begin{array}{l}\underset{t_d}{\overset{t}{{\displaystyle \int }}}{C}_{wb}\left(\tau -t_d\right)d\tau -\\ \underset{t_d+T_c}{\overset{t}{{\displaystyle \int }}}{C}_{wb}\left(\tau -t_d-T_c\right)d\tau \end{array}\right]+\\ C_p\left(t-t_d\right)\otimes \left(G{e}^{-\alpha \left(t-T_c\right)}+H{e}^{-\beta \left(t-T_c\right)}\right).\end{array}$$ (10)

The F2TC model has been applied in non-small cell lung cancer ²¹ and prostate cancer ²² using standard-temporal resolution dynamic PET, but has not been validated for high-temporal resolution brain data or BBB PS and CBF estimation, which is the focus of this work.

Parameter estimation

Model parameters were estimated by non-linear least squares fitting implemented using the lmfit Python library. ²³ Time delay $t_d$ was included as a jointly estimated parameter in non-linear least squares fitting together with all other model parameters. The S2TC model fitted for $v_b,$ $K_1,$ $k_2,$ $k_3,$ $k_4$ and $t_d$ (n = 6 parameters) whereas the F2TC model implementation fitted for $CBF,$ $T_c,$ $E,$ $k_2,$ $k_3,$ $k_4$ and $t_d$ (n = 7 parameters), noting $K_1\equiv CBF\cdot E.$ For S1TC and AATH model parameter estimation, we fixed $k_3=0$ and $k_4=0$ in the S2TC and F2TC fitting functions, respectively. Bounds and initialization parameters are described in Supplementary Table 1. Uniform weights were used across time points to compute residuals. Although it is possible to directly estimate PS as an F2TC model parameter, we did not find that this improved model fitting/parameter estimation and fitting for E simplified the constraint $0\le E\le 1.$ [¹⁸F]FDG PS was estimated using equation (3) with maximal E values restricted to 0.9999 to avoid indeterminate values in the logarithm. For [¹¹C]butanol, after comparison of S1TC and AATH models, we used S1TC K₁ as our reference estimate of CBF. ¹⁵ As detailed in the previous section, no blood-to-plasma correction was applied for [¹¹C]butanol ²⁰ and a population-based correction was applied for [¹⁸F]FDG using an exponential function described by Naganawa et al. ¹⁹

Image analysis

Our analysis focussed on cortical grey matter, subcortical grey matter, white matter, brainstem and whole cerebellum. To obtain time-activity curves of these regions, we used a deep learning-based tool ²⁴ to segment the Hammersmith brain atlas ²⁵ from static [¹⁸F]FDG PET images, separated the grey and white matter using an Otsu threshold, ²⁶ and re-grouped the Hammersmith regions to form the five brain regions of interest. An example brain segmentation and average tissue volumes across our analysis cohort are shown in Supplementary Figure 1. [¹¹C]butanol time-activity curves were obtained by co-registering and resampling the [¹⁸F]FDG PET image and brain segmentations using a rigid transformation computed by the ANTs software library. ²⁷ The ascending aorta was manually delineated to obtain a high-temporal resolution IDIF. For dynamic [¹⁸F]FDG PET frames beyond 5 min post injection, the FALCON ²⁸ package was used for inter-frame co-registration with respect to the 240–300 s frame.

For visualization, we performed voxel-wise kinetic analysis to generate quantitative maps of CBF, [¹⁸F]FDG K₁, [¹⁸F]FDG PS and [¹⁸F]FDG net uptake rate $(K_i=K_1{k}_3/(k_2+k_3),$ proportional to the cerebral metabolic rate of glucose). CBF maps were generated with both [¹¹C]butanol and [¹⁸F]FDG and BBB PS was calculated using both dual-tracer and single-tracer methods. Parametric images were generated from dynamic images reconstructed with 2.3-mm isotropic voxel size. Non-local means smoothing was applied post reconstruction via the kernel method. ²⁹ The kernel matrix was calculated using summed PET images ([¹⁸F]FDG: 0–5, 5–20, 20–40, 40–60 min; [¹¹C]butanol: 0–1, 1–2, 2–3 min) and 49 nearest neighbours within a 9 × 9 × 9 window.

Statistical analysis

The primary goal was to evaluate the quantitative agreement in [¹⁸F]FDG PS estimates between the conventional dual-tracer method and our single-tracer method. Because PS is derived from [¹⁸F]FDG K₁ and CBF using equation (3), we first investigated (1) how temporal resolution, analysis time window and model selection affect K₁ estimates and (2) the quantitative accuracy of [¹⁸F]FDG CBF estimates with respect to [¹¹C]butanol. For consistency, high-temporal resolution data were used for the main comparison of dual-tracer versus single-tracer PS estimates. Fitting quality was evaluated across all investigated models (S1TC, S2TC, AATH, F2TC; reversible and irreversible configurations) using the Akaike Information Criteria (AIC), ³⁰ which quantifies fitting error with adjustment for the number of model parameters. The model used for [¹⁸F]FDG K₁ in the dual-tracer method was selected based on the AIC. Linear correlation and quantitative agreement were evaluated by the Pearson coefficient and Bland–Altman analysis, ³¹ respectively. Mean percent differences and standard deviations were computed between parameter estimates to quantify systematic differences and variance across methods.

Results

Impact of high temporal resolution and model selection on [¹⁸F]FDG kinetics

We first investigated how temporal resolution, analysis time window and choice of kinetic model affect estimates of [¹⁸F]FDG K₁ (Figure 1). For 60-min time-activity curves, the AIC preferred reversible two-tissue compartment models over irreversible models, and the S1TC and AATH models could not adequately fit 60-min [¹⁸F]FDG data. At high temporal resolution (1-s early frames), the AIC preferred the F2TC model over the S2TC model, but not for standard temporal resolution (10 s early frames). AIC results and model fittings for high-temporal resolution 60-min time-activity curves are shown in Figure 1; additional AIC comparisons at standard temporal resolution and for high-temporal resolution 2-min time-activity curves are shown in Supplementary Figure 2.

Figure 1.

Model selection and time-activity curve fitting for [¹⁸F]FDG: (a) Akaike Information Criteria for high-temporal resolution 60-min time-activity curve fitting with the S1TC and S2TC models and AATH model and the F2TC model (i and r). Note, the order of the bar plots correspond to the order of the legend entries. (b, c) Examples of high-temporal resolution time-activity curve fits in cortical grey matter for the S2TC and F2TC models for (b) irreversible and (c) reversible configurations. The reversible configurations demonstrated improved fits.

S1TC: standard one-tissue compartment; S2TC: standard two-tissue compartment; AATH: adiabatic approximation to the tissue homogeneity; F2TC: flow-modified two-tissue compartment; i: irreversible; r: reversible.

Figure 2 shows that [¹⁸F]FDG K₁ estimates are affected by the temporal resolution and model selection. K₁ estimates for high-temporal resolution data were larger than those from standard temporal resolution data even when using the same kinetic model. F2TC K₁ estimates were smaller than those from the S2TC model even when using the same high-temporal resolution time-activity curves. All combinations (temporal resolution, model selection, analysis duration) of K₁ estimates were strongly correlated (Pearson coefficient >0.95), but between configurations, mean differences were as high as ≈36% (60-min standard temporal resolution irreversible S2TC vs 2-min high-temporal resolution S1TC) in cortical grey matter.

Figure 2.

Cortical grey matter [¹⁸F]FDG K₁ estimates (n = 18 samples/method) with respect to temporal resolution and kinetic model: (a) comparisons of K₁ estimates between kinetic models, temporal resolution and analysis time windows, (b) correlation matrix for K₁ estimates across all methodological configurations, and (c) matrix of mean percent differences for K₁ estimates across methodological configurations. A negative difference indicates the row value was smaller than the column value.

S2TC: standard two-tissue compartment model; F2TC: flow-modified two-tissue compartment model; AATH: adiabatic approximation to the tissue homogeneity model; i: irreversible model ${(k}_4=0)$ ; r: reversible model.

[¹⁸F]FDG kinetic parameter estimates using two-tissue compartment models and high-temporal resolution 60-min data are shown in Supplementary Tables 2 and 3. [¹⁸F]FDG K₁ to k₃ estimates derived from the irreversible S2TC and F2TC models were comparable to values reported by Larsson et al. ¹¹ and Sari et al. ³² who used irreversible S2TC modelling and high-temporal resolution dynamic [¹⁸F]FDG imaging with long axial-field-of-view PET. Using reversible S2TC or F2TC models, we found that [¹⁸F]FDG k₃ was approximately one order-of-magnitude larger than [¹⁸F]FDG k₄, in agreement with a classical study by Huang et al. ³³ Differences between irreversible and reversible [¹⁸F]FDG micro kinetics were generally larger than those between S2TC and F2TC models. AIC analysis considerably favoured the reversible models.

[¹¹C]butanol versus [¹⁸F]FDG estimation of CBF

Next, we validated the quantification of CBF derived from [¹⁸F]FDG versus those from [¹¹C]butanol (Figure 3). For [¹¹C]butanol, we used S1TC K₁ as the reference CBF measure because the AIC was very similar between the AATH and S1TC model fits (Supplementary Table 4). This is likely because [¹¹C]butanol E ≈ 100% (i.e. CBF ≈ K₁) and thus the extra parameter in the AATH model may not improve fitting. Nonetheless, AATH CBF and K₁ both agreed with S1TC [¹¹C]butanol K₁ values.

Figure 3.

Correlation (left column) and Bland–Altman (right column) plots comparing CBF estimates from [¹¹C]butanol and [¹⁸F]FDG. For the top row, [¹⁸F]FDG CBF was derived from 2-min high-temporal resolution time-activity curves and the AATH model. The bottom row derived [¹⁸F]FDG CBF using 60-min high-temporal resolution data and the F2TCr model.

CBF: cerebral blood flow; AATH: adiabatic approximation to the tissue homogeneity; F2TCr: flow-modified two-tissue compartment reversible; GM: grey matter.

[¹¹C]butanol CBF values were strongly correlated with [¹⁸F]FDG CBF estimates using both the AATH model (0–2-min data) and the proposed F2TC model (full 60-min data; Figure 3) on high-temporal resolution time-activity curves. [¹⁸F]FDG CBF estimates underestimated that of [¹¹C]butanol by ≈15% on average. Mean differences effectively reduced to zero when [¹⁸F]FDG CBF estimates were scaled up by ≈15% (Supplementary Figure 3). Nonetheless, original [¹⁸F]FDG CBF estimates were used for single-tracer PS calculations for consistency. The standard deviation of differences relative to [¹¹C]butanol was ±0.08 mL/min/cm³ for both AATH and F2TC models, and by Bland–Altman analysis, 95% of measurements fell within ±0.16 mL/min/cm³ of the mean difference.

Dual-tracer versus single-tracer estimation of [¹⁸F]FDG BBB PS and E

Given the observed differences in K₁ estimates across temporal resolution and model selection, dual-tracer PS was computed using K₁ of the reversible F2TC model, which yielded the lowest AIC. This also ensured that high temporal resolution data were used consistently with the single-tracer method. With this configuration, single-tracer and dual-tracer PS estimates had excellent correlation and agreement (Figure 4), with percent mean ± standard deviation differences of 6.7% ± 5.8% for the AATH model and 3.7% ± 4.2% for the F2TC model. Comparisons of single-tracer PS estimates against dual-tracer PS estimates derived from the S2TC model at standard and high temporal resolution (Supplementary Figures 4 and 5) also showed excellent correlation albeit with greater bias due to methodological differences in K₁ estimates. When calculating single-tracer PS estimates with [¹⁸F]FDG CBF values scaled up by 15% (to reduce bias against [¹¹C]butanol CBF), the bias in PS estimates became nearly negligible (Supplementary Figure 6). Agreement in [¹⁸F]FDG extraction fraction between dual-tracer and single-tracer methods was weaker but still significant (Figure 5).

Figure 4.

Correlation (left column) and Bland–Altman (right column) plots comparing [¹⁸F]FDG PS estimates from the dual-tracer and proposed single-tracer method. The top row derived [¹⁸F]FDG PS from 2-min high-temporal resolution time-activity curves and the AATH model. The bottom row derived [¹⁸F]FDG CBF using 60-min high-temporal resolution data and the F2TCr model. Dual-tracer PS was calculated using [¹¹C]butanol CBF and [¹⁸F]FDG K₁ derived from 60-min high-temporal resolution data and the F2TCr model.

PS: permeability-surface area; AATH: adiabatic approximation to the tissue homogeneity; F2TCr: flow-modified two-tissue compartment reversible; GM: grey matter.

Figure 5.

Correlation (left column) and Bland–Altman (right column) plots comparing [¹⁸F]FDG extraction fraction (E) estimates from the dual-tracer and proposed single-tracer method.

GM: grey matter; AATH: adiabatic approximation to the tissue homogeneity; F2TCr: flow-modified two-tissue compartment reversible.

Single-tracer multiparametric brain PET imaging

Figure 6 illustrates the potential for single-tracer multiparametric brain imaging with high-temporal resolution dynamic [¹⁸F]FDG PET. Voxel-wise parametric imaging results largely agreed with our regional analysis, with single-tracer [¹⁸F]FDG BBB PS images appearing very similar to those from the dual-tracer method. [¹⁸F]FDG and [¹¹C]butanol CBF maps were overall similar except at highly vascularized regions like the middle cerebral artery and sagittal sinus. Such discrepancies were mostly localized to voxels with high vascular volume fraction and are consistent with the interpretation that [¹¹C]butanol CBF is weighted towards capillary flow whereas [¹⁸F]FDG CBF reflects the intravascular flow rate through the entire regional vascular volume. ¹³ [¹⁸F]FDG K₁ and BBB PS appeared spatially similar but differed in scale. Unlike the AATH model, the F2TC model allows simultaneous quantification of [¹⁸F]FDG K_i as the full 60-min time-activity curve can be appropriately fitted. Parametric images without non-local means smoothing are shown in Supplementary Figure 7.

Figure 6.

Multiparametric PET imaging of blood–brain barrier PS, CBF, [¹⁸F]FDG K₁ and [¹⁸F]FDG K_i. Images on the top row were generated with only [¹⁸F]FDG data. Dual-tracer [¹⁸F]FDG PS (bottom left) was generated using [¹⁸F]FDG K₁ and [¹¹C]butanol CBF (bottom right). Note the prevalence of large arteries (e.g. middle cerebral and anterior cerebral artery) and veins (e.g. sagittal sinus and straight sinus) in the [¹⁸F]FDG CBF map but not in the [¹¹C]butanol CBF map. Concordance with [¹¹C]butanol CBF improves where the large vessel volume fraction is low.

PS: permeability-surface area; CBF: cerebral blood flow.

Discussion

Single-tracer approaches for effectively assessing molecular BBB permeability were previously not possible with conventional PET systems due to their inadequate sensitivity and thus required serial dual-tracer scans to separately measure CBF and K₁. Here, we validated the quantification of [¹⁸F]FDG BBB PS derived from a single-tracer method that leverages a high-sensitivity PET system by comparing against the conventional dual-tracer method in human subjects. We confirmed our prior work ¹² that with high-temporal resolution imaging of just the first 2 min, our single-tracer method allows quantification of BBB PS within a small margin of error (6.7% ± 5.8%) compared to the conventional dual-tracer approach. This bias was further reduced (3.7% ± 4.2%) when using the full 60-min data for which we validated high-temporal resolution kinetic modelling methods with comparisons against standard compartment models. Quantification of BBB PS was enabled by accurate estimation of CBF with [¹⁸F]FDG using either 2- or 60-min data as validated against the reference flow tracer [¹¹C]butanol. By mitigating the need for complex dual-tracer procedures, our method may broaden opportunities for future quantitative studies of diverse transport systems and molecular mechanisms at the BBB, advancing our understanding of their roles in neurological and systemic disorders.

The availability of high-temporal resolution PET time-activity curves prompts examining whether our conventional kinetic models remain appropriate. In our earlier work,^12,13 we showed that AIC favoured the AATH model over the standard compartment model for high-temporal resolution early time-activity curves. However, because the AATH model effectively comprises a single tissue compartment, we chose an empirical 2-min analysis window to minimize the signal contribution of phosphorylation kinetics. To overcome this limitation, we validated the F2TC model for full 60-min dynamic data and demonstrated that high-temporal resolution 60-min time-activity curves require more complex models such as the F2TC rather than the conventional S2TC model. Estimating the additional parameter in the F2TC model was supported by the higher temporal resolution and greater number of data points (1 s/frame over the first 1 min). While our study used total-body PET, our method is compatible with other high-sensitivity PET systems like the NeuroEXPLORER, which supports high-temporal resolution imaging and non-invasive measurement of the arterial input function from the carotid arteries. ³⁴

Our study agrees with prior work suggesting that K₁ estimates are affected by several technical factors such as temporal resolution, analysis time window and choice of kinetic model. ³⁵ Although K₁ estimates across all tested configurations had excellent correlation (R > 0.95), mean differences of up to ≈36% were observed. High temporal resolution alone increased K₁ estimates compared to standard temporal resolution, even when using the same model, suggesting that temporal resolution itself causes a systematic shift. The choice of kinetic model also contributed differences as F2TC K₁ estimates were systematically smaller than those from the S2TC model for the same high-temporal resolution time-activity curve. To maintain consistency, we derived dual-tracer PS using high-temporal resolution data and F2TC modelling for [¹⁸F]FDG. Whether F2TC-derived K₁ values are ‘more correct’ is unclear; instead, we strived for methodological consistency, and results using the standard model are provided in the Supplementary Materials. One approach for experimentally validating the true value of K₁ (CBF × E) is to use arteriovenous sampling and the Fick Principle to independently measure whole-brain CBF and [¹⁸F]FDG extraction fraction. This may be possible using high-resolution brain PET systems like the NeuroEXPLORER. ³⁶

Our study found that reversible two-tissue compartment models (k₄ ≠ 0) were favoured over irreversible models (k₄ = 0) of [¹⁸F]FDG kinetics based on AIC analysis. The cerebral dephosphorylation rate of [¹⁸F]FDG is very slow^33,37 (k₃/k₄ ≈ 10; also found in our analysis in Supplementary Tables 2 and 3) and the estimation of a non-zero k₄ has been attributed to tissue heterogeneity. ³⁸ As such, irreversible models are commonly employed for [¹⁸F]FDG brain kinetic modelling to improve parameter identifiability of the remaining microkinetic parameters. Although further validation is required, the higher sensitivity and spatial resolution performance of modern PET scanners, such as the uEXPLORER total-body PET system, may allow more reliable quantification of microkinetic parameters like the phosphorylation and dephosporylation rates of [¹⁸F]FDG. ³⁹ Alternatively, we have found that high-temporal resolution imaging may support modelling the interstitial space via a three-tissue irreversible compartment model. ⁴⁰ Further work is required to reconcile vascular versus interstitial transport and how their added modelling affects the interpretation and reliability of tissue microkinetic parameters.

This study builds on our pilot validation ¹³ of [¹⁸F]FDG-derived blood flow against [¹¹C]butanol by studying a larger cohort with a dedicated focus on the brain. Using high-temporal resolution data, [¹⁸F]FDG CBF estimates were strongly correlated (R ≈ 0.85) with [¹¹C]butanol CBF using both the AATH model (2-min early data) and the F2TC model (full 60-min data). Although [¹⁸F]FDG CBF estimates were systematically ≈15% less than that of [¹¹C]butanol, the standard deviation of differences remained relatively small (±0.08 mL/min/cm³). The source of this systematic bias is still unclear; however, one avenue for further investigation involves the arterial input function. While the blood–plasma distribution of [¹⁸F]FDG is well-characterized for minute-to-hour timescales,^19,37 its early-phase blood distribution is less understood due to the challenge of sampling arterial blood at the 1–2 s temporal resolution used here. Alternatively, internal dispersion from the aorta to the brain may have had a greater impact on [¹⁸F]FDG than [¹¹C]butanol due to the former’s lower extraction fraction. Lastly, [¹¹C]butanol CBF is weighted towards capillary flow whereas [¹⁸F]FDG CBF reflects total intravascular flow. This conceptual difference may contribute to bias though the impact at the regional level may be limited, as supported by the near-100% [¹¹C]butanol extraction fraction estimated with the AATH model (Supplementary Table 4). ¹² Nonetheless, we show that [¹⁸F]FDG CBF correlates well with [¹¹C]butanol CBF, potentially obviating the need for a challenging flow tracer scan.

Altogether, our proposed high-temporal resolution dynamic PET method provides an effective single-tracer approach for estimating [¹⁸F]FDG BBB PS. Single-tracer PS estimates, using either 2 or 60 min of data, had excellent agreement with the dual-tracer method, especially when using consistent temporal resolutions and kinetic models. We also showed that dual-tracer PS estimates made with the S2TC model may be systematically shifted due to methodological differences, which may inform comparisons against other studies using the standard compartment model. While [¹⁸F]FDG extraction fraction estimates showed weaker agreement, we consider E a method parameter for calculating PS, and PS was nonetheless accurate. One explanation for the weaker agreement in E is that CBF and E have high negative covariance, ¹² thus reducing the variance of PS estimates. We have also extensively studied the practical identifiability of AATH model parameters in the past and found that PS, K₁ and CBF were exceptionally identifiable for the kinetics expected in the brain. ¹²

The favourable accuracy of parameter estimates with the 2-min protocol may facilitate clinical integration through dual-time point imaging, in which CBF, [¹⁸F]FDG K₁ and [¹⁸F]FDG PS are derived from an early dynamic scan and cerebral glucose metabolism is assessed using a standard delayed static acquisition or the relative Patlak plot method. ⁴¹ Alternatively, for more comprehensive characterization, a full 60-min dynamic scan can allow quantification of [¹⁸F]FDG K_i and other microkinetic parameters, as well as reduce bias in K₁ and PS estimates. Together, these approaches expand opportunities for quantitative multiparametric brain PET imaging, allowing simultaneous assessment of BBB permeability, CBF, glucose metabolism and microkinetic parameters like K₁ all from a single, widely used [¹⁸F]FDG tracer. While our analysis here focussed on the brain, total-body PET allows dynamic imaging of the entire human body, and future work will examine modelling approaches for quantification of extracerebral organs for total-body multiparametric imaging.

As the purpose of this work was to validate consistency between single-tracer and dual-tracer PS estimates, we did not account for several physiological factors relevant for absolute PS quantification. Standards for DCE-MRI based on extracellular gadolinium-based contrast agents have been proposed, ⁴² but may differ for molecular PET tracers. For example, we did not convert whole-blood CBF to plasma flow when calculating PS. In DCE–MRI, plasma flow is obtained by scaling whole-blood flow by 1 − haematocrit (Hct; often assumed ≈0.45), but this scaling factor may differ for [¹⁸F]FDG which can partially distribute in erythrocytes. ³⁷ Physiological corrections may therefore depend on each tracer and deriving tracer-specific corrections was outside the scope of this work. The absence of such corrections may have contributed to the underestimation of CBF values with [¹⁸F]FDG observed in this work. Furthermore, we expect our PS estimates to partially include interstitial-to-intracellular transport, though this challenge is not exclusive to our method. ⁴³ More advanced models, such as a three-tissue compartment model, ⁴⁰ may potentially help disentangle these components, though further validation is required.

Here, we used [¹⁸F]FDG, the most widely used PET tracer, as a representative example to demonstrate that our method can better disentangle CBF and PS. These results are likely applicable to other tracers, broadening opportunities to study molecular BBB permeability. Our method would benefit from further in vivo validation in tracers with higher extraction fractions for which K₁ is less coupled with PS, although its robustness across a wide range of extraction fractions was previously characterized in simulation studies. ¹² While [¹⁸F]FDG K₁ and PS are similar in magnitude, PS more specifically describes a property of the BBB, whereas K₁ reflects a mixture of CBF and PS. Many neuroreceptor and protein-targeted PET tracers also have non-negligible radiometabolites, necessitating corrections with arterial blood sampling or population-based parent fraction curves. Our original method, which estimates BBB PS using the first 2 min of dynamic data, may help mitigate the impact of radiometabolites. ¹² Further in vivo studies to determine how tracer-specific properties, such as extraction fraction, radiometabolites and binding/retention kinetics, affect the quantification of CBF, K₁ and PS, will help establish the broader generalizability of our single-tracer framework beyond [¹⁸F]FDG.

This study had limitations. First, cerebral haemodynamics may have changed in the interval between [¹¹C]butanol and [¹⁸F]FDG scans. While most participants were scanned within ~2.5 h, test-retest studies and shorter-interval comparisons (e.g. using [¹⁵O]water) may be informative. Second, we did not specifically validate the quantification of BBB PS with a known permeability defect or an otherwise known change in [¹⁸F]FDG BBB permeability. However, we previously showed that our [¹⁸F]FDG BBB permeability estimates were sensitive to blood glucose differences across participants consistent with competitive transporter effects, and that CBF and [¹⁸F]FDG PS may become decoupled with severe metabolic dysfunction-associated steatoheptatis. ¹² Third, validation was only performed with [¹⁸F]FDG; further studies are needed for other radiotracers. Lastly, we analyzed regional time-activity curves with high signal-to-noise ratio owing to the high sensitivity of total-body PET and the relatively large region of interests used in this work (Supplementary Figure 1). Future work will focus on higher noise levels expected for voxel-wise parametric imaging, smaller brain regions, lower injected activities and lower-sensitivity PET systems.

In conclusion, high-temporal resolution dynamic PET and distributed kinetic modelling enables single-tracer quantification of molecular BBB permeability and CBF using [¹⁸F]FDG. We examined methodological factors influencing the estimation of BBB PS, including temporal resolution and model selection for estimating [¹⁸F]FDG K₁, and validated [¹⁸F]FDG-derived CBF against [¹¹C]butanol. Our proposed method may serve as an effective quantitative tool for assessing BBB permeability across a wide range of molecular pathways and neurological and systemic diseases.