Sage Journals: Discover world-class research

Abstract

Positron emission tomography (PET) with ¹⁵O-labeled water can provide reliable measurement of cerebral blood flow (CBF). Quantification of CBF requires knowledge of the arterial input function (AIF), which is usually provided by arterial blood sampling. However, arterial sampling is invasive. Moreover, the blood generally is sampled at the wrist, which does not perfectly represent the AIF of the brain, because of the effects of delay and dispersion. We developed and validated a new noninvasive method to obtain the AIF directly by PET imaging of the internal carotid artery in a region of interest (ROI) defined by coregistered high-resolution magnetic resonance angiography. An ROI centered at the petrous portion of the internal carotid artery was defined, and the AIF was estimated simultaneously with whole brain blood flow. The image-derived AIF (IDAIF) method was validated against conventional arterial sampling. The IDAIF generated highly reproducible CBF estimations, generally in good agreement with the conventional technique.

INTRODUCTION

Quantification of cerebral blood flow (CBF) is important for the understanding of normal and pathologic brain physiology. While several imaging techniques for measuring CBF have been described,¹ positron emission tomography (PET) with [¹⁵O]H₂O^2,3 is the current gold standard,⁴ especially when absolute measures are of interest. [¹⁵O]H₂O-PET is commonly used in conjunction with other ¹⁵O tracers, i.e., [¹⁵O]CO and [¹⁵O]O₂, to measure cerebral blood volume⁵ and cerebral oxygen metabolic rate.^2,6 To quantify these parameters, mathematical modeling is required, for which knowledge of the arterial input function (AIF) is a prerequisite.^2,3 Currently, arterial blood sampling is the standard approach for obtaining the AIF; however, this is invasive and technically challenging. Therefore, investigation of other approaches for estimating the AIF is warranted.

Several approaches have been developed to estimate the AIF without arterial blood sampling. One approach calibrates a population template AIF to each specific scan,^7,8 but this approach neglects individual differences in physiology and uncontrolled timing variability of tracer introduction. A second approach applies partial volume correction to region of interest (ROI) time-activity curves (TACs) obtained from imaging major blood pools, such as large arteries and the cardiac ventricles.^9,10 An approach based on factor analysis has also been developed and validated for AIF estimation;^11,12 a difficulty with this approach is that the linear system does not have a unique solution and additional information, such as radioactivity measurements in a few blood samples, is generally required to calibrate the estimated AIF.^11,13 Simultaneous estimation of both AIF and tracer kinetic parameters has been proposed but this requires full kinetic modeling and assuming a particular mathematical model of the AIF.^14–16 Although considerable effort has been devoted to the development of image-based AIF estimation, a reliable method is still unavailable. In fact, a recent study evaluating eight methods for quantitative estimation of the cerebral metabolic rate of glucose utilization¹⁷ found that none reliably estimates the AIF or cerebral metabolic rate of glucose utilization in human brain without the requirement for taking blood samples.

Here, we present a noninvasive image-derived AIF (IDAIF) method that uses coregistered high-resolution magnetic resonance angiography (MRA), which identifies the precise location of the carotid artery. A joint estimation approach is used to estimate the AIF and CBF. The proposed method was tested for reproducibility and also validated against the conventional technique based on arterial blood sampling.

MATERIALS AND METHODS

Human Subjects

Imaging data were acquired in healthy volunteer subjects without history of neurologic disorders. All procedures were approved by the Human Studies and the Radioactive Drug Research Committees at our institution. Subjects provided written informed consent and were paid for their participation. Of these subjects, a first group of eight participants (age 29 ± 6 years, 4 females) was recruited to test the reproducibility of the CBF measurements as well as to compare two different PET scanners. Each subject in this group underwent four repeated [¹⁵O]H₂O scans in a single session on a Siemens EXACT 962 HR + scanner (Siemens Medical Solutions USA, Inc., Mallvern, PA, USA), and another four ¹⁵O-water scans in a single session on a Siemens Biograph 40 PET/CT scanner (Siemens Medical Solutions USA, Inc.). Each [¹⁵O]H₂O scan started with 35 2-second frames, followed by six 5-second frames and eight 10-second frames. The PET images were reconstructed using filtered-back projection method without smoothing. A second group of 12 participants (age 30 ± 7 years, 6 females) completed the study to validate the IDAIF method against the conventional arterial sampling AIF (ASAIF) technique. Each subject in this group underwent one [¹⁵O]H₂O scan on the Siemens EXACT 962 HR+ scanner with frame setup similar to group 1. During these PET studies, arterial blood was withdrawn at 5 mL/min from the radial artery through narrow bore tubing to a lead shielded scintillation detector that measures positron emissions with 1 second temporal resolution.¹⁸ Blood sampling started simultaneously with time-binned PET scanning and continued for 2 min. Each participant in both groups also underwent MRI using a Siemens Trio 3T scanner (Siemens Medical Solutions USA, Inc.), which included one Magnetization Prepared Rapid Gradient Echo (MPRAGE) scan and one time-of-flight MRA scan.

Image Processing

Cross-modal registration, i.e., MPRAGE to PET and MRA to MPRAGE, was achieved using the vector gradient method.¹⁹ Within modality registration, e.g., PET to PET, was achieved using standard techniques²⁰ implemented with an in-house software. Atlas registration (12-parameter affine) was performed via the MPRAGE against a template representing young healthy volunteers using the same software. A modified adaptive segmentation algorithm²¹ was used to automatically segment the MRA images into three tissue types, (1) arteries (Figure 1A); (2) signal contributing tissue, i.e., brain tissue, muscle, etc.; and (3) nonsignal contributing structures, i.e., bone, cerebrospinal fluid, and airways. Three versions of arterial ROIs (ROIa) were created for AIF estimation. The first version (ROIa1) was centered on the left and right petrous segments of the internal carotid arteries, which were manually defined based on the automatically segmented arteries (Figure 1B). The second version (ROIa2) used all the arterial regions in a rectangular box (7 cm × 9 cm × 3.5 cm) defined in the atlas space centered at the petrous portion of the internal carotid artery (Figure 1C). The third version (ROIa3) used all arterial regions within the time-of-flight MRA field of view, less a 1.5-cm border to minimize the boundary effects (Figure 1D). FreeSurfer v5.1 (Martinos Center for Biomedical Imaging, Charlestown, MA, USA) was used to segment the MPRAGE image. Three additional ROIs were defined based on the FreeSurfer segmentation: a whole-brain ROI (ROI1), a cortical gray matter ROI (ROI2), and a deep white matter ROI (ROI3) (Figure 2) to facilitate IDAIF and CBF estimation as described below.

Figure 1.

Three-dimensional rendering of segmented artery from the time-of-flight magnetic resonance angiography images (A); the arterial region of interest (ROI) version one (ROIa1) (B); the arterial ROI version two (ROIa2) (C); and the arterial ROI version three (ROIa3) (D).

Figure 2.

Example Magnetization Prepared Rapid Gradient Echo (MPRAGE) images (A); whole-brain region of interest (ROI) (B); cortical gray matter ROI (C); and deep white matter ROI (D). The three ROIs (B, C, D) correspond to the regions needed for the Image-Derived Arterial Input Function (IDAIF) algorithm as described in Materials and methods.

Image-Derived Arterial Input Function Method

The primary difficulty in PET image-based AIF estimation is limited spatial resolution. The diameter of the arteries feeding the brain is comparable to or smaller than the spatial resolution of PET (nominally, 6 mm).²² Simple ROI-based measurement greatly underestimates the peak portion of the AIF because of partial volume effects. To account for limited spatial resolution, the PET data were modeled as a linear mixture of the true arterial signal and a surrounding tissue component (Equation 1).

C_{ROI} = rCa + sCt,

(1)

in which, C_ROI is the measured ROIa activity concentration, Ca is the estimated arterial activity concentration (the AIF), and Ct is the surrounding tissue activity concentration. Ct was modeled using Equation (2):²³

Ct = rCBF \times Ca \otimes e^{- rCBF \times t / λ},

(2)

where rCBF is regional cerebral blood flow and ⊗ denotes the convolution operation. The parameter, λ, is the brain-blood partition coefficient for water, for which a fixed value of 0.90 was used throughout this study.²⁴ The variables, r and s in Equation (1), are parameters accounting for partial volume effects, determined for each ROIa (Figures 1B and 1D) using an empirically derived point spread function. This point spread function (full width at half maximum (FWHM) = 8.3 mm) was set on the basis of a single [¹⁵O]H₂O CBF study (separate from data acquired in this study) and calibrated so that the IDAIF and the ASAIF methods generated the same mean whole-brain CBF.

To estimate the AIF, we adopted a simultaneous estimation approach, in which three additional ROIs were defined to obtain three distinct TACs. The AIF (Ca) was estimated by minimizing the following cost function:

\begin{matrix} Q (rCBF1, rCBF2, rCBF3, rCBFa) = w_{1} \sum_{i = 1}^{K} {[C_{ROIa}^{PET} (i) - C_{ROI3}^{MOD} (i) - C_{ROIa}^{MOD}]}^{2} \\ + w_{2} \sum_{i = 1}^{K} {[C_{ROI2}^{PET} (i) - C_{ROI2}^{MOD} (i)]}^{2} + w_{3} \sum_{i = 1}^{K} {[C_{ROI3}^{PET} (i) - C_{ROI3}^{MOD} (i)]}^{2} \end{matrix}

where i is the frame index, K is the total number of frames within the 40 second window, MOD refers to model estimated TACs, PET refers to PET measured TACs, and w₁, w₂, w₃ are weighting factors proportional to the square root of ROI volumes. The parameters, rCBF1, rCBF2, and rCBF3, are the estimated regional blood flow in the whole-brain ROI (ROI1, Figure 2B), the cortical gray matter ROI (ROI2, Figure 2C), and the deep white matter ROI (ROI3, Figure 2D). The parameter, rCBFa, is the regional blood flow for the vicinity of the selected ROIa (Figure 1). A whole-brain ROI was chosen as the basis for AIF derivation because this maximizes counting statistics, and hence, the signal-to-noise ratio. The other two ROIs (cortical gray matter and deep white matter) were chosen for their difference in kinetics to improve parameter estimation. Other ROI1 definition is possible although not tested within this study. Given rCBF1, rCBF2, rCBF3, and rCBFa, the cost function, Q, can be computed by first calculating the AIF (Ca) based on the following equation:

Ca = \frac{1}{rCBF1} \cdot \frac{{dC}_{ROI1}}{dt} + \frac{1}{λ} C_{ROI1},

(4)

derived by rearrangement of the blood flow compartmental model:³

\frac{dCt}{dt} = rCBF \times Ca - \frac{rCBF}{λ} Ct

(5)

Then, model-based regional TACs, C_ROI2^MOD and C_ROI3^MOD, can be calculated based on the estimated AIF and rCBFs (rCBF2 and rCBF3) according to Equation (2); C_ROIa^MOD can be calculated according to equations (1) and (2), using the estimated AIF and rCBFa. Finally, Q can be computed using Equation (3) from the model-based regional TACs, i.e., C_ROIa^MOD, C_ROI2^MOD, and C_ROI3^MOD, and the PET measured regional TACs, i.e., C_ROI2^PET, C_ROI3^PET, and C_ROIa^PET. After the optimal set of rCBFs is determined by minimization of the cost function Q, the final AIF can be calculated according to Equation (4). To compare IDAIF versus ASAIF, the cost function was evaluated over a 40-second window starting at the onset of brain activity, for reasons discussed in the next section. Whole-brain CBF (rCBF1), estimated during the optimization process, was used to evaluate the reproducibility of the IDAIF technique and to compare IDAIF versus the conventional ASAIF technique. In addition, regional blood flow was also evaluated for visual cortex gray matter based on the estimated AIF using both approaches for comparison.

Cerebral Blood Flow Calculation with Arterial Sampling (Arterial Sampling Arterial Input Function)

Estimation of CBF on the basis of arterial sampling currently is the gold standard. Nevertheless, this technique is systematically biased by delay and dispersion caused by both the finite length of tubing between the wrist and the activity detector and hemodynamic differences between the radial artery and the arterial blood supplying the brain.¹⁸ A modeling approach was adopted to estimate delay, dispersion, and rCBF simultaneously within the whole-brain ROI based on the following equation:^25–27

Ca (t) = g (t + Δ t) + τ \frac{d g (t + Δ t)}{d t}

(6)

where g(t) is the activity profile measured using the arterial sampling device, τ is the dispersion factor, and Δt is the delay time. Three parameters, rCBF, τ, and Δt, were estimated simultaneously so that the whole-brain TAC measured by PET optimally matched, in the least squares sense, the whole-brain TAC estimated using equations (2) and (6) . Again, a 40-second window was used for model fitting. The duration of this window was constrained mainly by the arterial sampling data, which was substantially delayed because of the distance (~60 cm) between the radial artery access point and the activity detector. The timing window was chosen so that a complete (40 seconds) delay and dispersion corrected AIF was available for all data sets.

Statistical Analysis

In experiment 1, in which each subject had four repeated [¹⁵O]H₂O scans, reproducibility of estimated hemodynamic parameters was evaluated as the within-subject, repeated measures standard deviation (SD_w) of whole-brain CBF. The within-subject coefficient of variation (CV_w%) was estimated as the percentage ratio of SD_w to the mean CBF for a particular subject.

To evaluate the agreement of CBF estimated with the ASAIF versus IDAIF, intraclass correlation (ICC) coefficients were calculated²⁸ in addition to Pearson's correlation. Absolute (Δ) and percentage differences (Δ%) were also evaluated in individuals according to the formula,

Δ = {CBF}_{ID} - {CBF}_{AS}

(7)

Δ % = \frac{Δ}{{CBF}_{AS}} \times 100 %

(8)

where ID refers to the CBF estimated using IDAIF and AS refers to the CBF estimated using arterial sampling data.

RESULTS

Study Group 1: Within-Subject Reproducibility and Comparison of Two Scanners

Example brain ROI TACs and the IDAIF are illustrated in Figure 3. Table 1 lists estimated whole-brain rCBF (rCBF1) based on data acquired on the HR+ scanner for the eight subjects in group 1. The mean within-subject coefficient of variation (CV_w) for rCBF1 estimated using the three alternative arterial ROIs (Figures 1B and 1D) was 7.83%, 6.38%, and 6.08%, respectively. This result shows that IDAIF-based CBF estimation is only moderately affected by the selected aROI. This property most likely reflects calculation of the AIF using the whole-brain TAC. The rCBF measurements obtained using the PET/CT scanner were comparable to those obtained using the HR+ scanner as assessed by intrasubject variability and mean value. The two sets of measurements were also highly correlated (r = 0.9553) (Figure 4).

Table 1

Whole brain cerebral blood flow (CBF) (mL/100 g per minute) estimation using image-derived arterial input function on the EXACT 962 HR+ scanner with three different versions of arterial ROI (ROIa)

	Mean CBF (mL/100 g per minute)			SDw (mL/100 g per minute)			CVw%
	ROIa1	ROIa2	ROIa3	ROIa1	ROIa2	ROIa3	ROIa1	ROIa2	ROIa3
1	39.23	41.61	41.41	1.59	1.78	0.86	4.04	4.29	2.07
2	28.03	32.85	32.44	5.26	5.07	5.01	18.76	15.45	15.43
3	31.56	36.63	39.39	0.93	1.62	1.62	2.96	4.42	4.10
4	30.19	28.39	25.92	4.31	2.41	2.30	14.28	8.49	8.89
5	30.56	32.10	29.75	1.19	2.19	1.16	3.88	6.83	3.88
6	36.35	38.03	35.58	2.01	1.05	0.60	5.53	2.75	1.69
7	34.98	37.15	35.61	1.84	1.85	2.19	5.25	4.99	6.16
8	53.14	53.98	59.41	4.20	2.07	3.80	7.91	3.83	6.39
Mean	35.50	37.59	37.44	2.67	2.26	2.19	7.83	6.38	6.08

ROIa1 has the least number of voxels while ROIa3 had the largest number of voxels. Within-subject standard deviation (SD_w) and within-subject coefficient of variation (CV_w%) were also reported.

Figure 3.

Example time-activity curves for whole brain (WBTAC), cortical gray matter (GMTAC), deep white matter (WMTAC), and estimated image-derived arterial input function (IDAIF). PET, positron emission tomography.

Figure 4.

Comparison of whole-brain cerebral blood flow (CBF) measurements on the EXACT 962 HR+ scanner and the Biograph 40 PET/CT scanner.

Study Group 2: Image-Derived Arterial Input Function Versus Conventional Arterial Sampling

Table 2 lists whole-brain CBF estimated in the same subjects, using both the arterial sampling technique and the image-based technique. An example IDAIF together with a delay and dispersion corrected ASAIF is illustrated in Figure 5. Among the 12 data sets analyzed, 2 of them (#11 and #12) were identified as outliers. Extremely low blood flow was observed in data set #11; we suspect that this anomaly is because of excessive patient motion. Data set #12 had unusually large dispersion factor and long delay time, indicating possible arterial sampling failure. Excluding these two outliers, the ICC between whole-brain CBF computed by ASAIF and IDAIF (arterial ROIa2) was 0.77. The Pearson's correlation between the two measurements was 0.89 (P = 0.0006) excluding the outliers. Mean whole-brain CBF estimated using the IDAIF technique was 2.67 mL/100 g per minute, or 7.66% greater than the result obtained using the ASAIF technique. The absolute and relative difference was calculated between IDAIF (ROIa2) and ASAIF. The visual cortex gray matter rCBF using the two approach was also highly correlated (r = 0.93; P = 0.0001; ICC = 0.82), the IDAIF-based rCBF was 4.26 mL/100 g per minute, or 9.51% greater than the result obtained using the ASAIF technique.

Table 2

Whole brain CBF (mL/100 g per minute) measurements obtained using the arterial sampling technique (ASAIF) and the image-derived AIF (IDAIF) technique with three different versions of arterial ROIs (ROIa1, ROIa2, and ROIa3) as well as the absolute (Δ) and relative difference (Δ%)

	ASAIF			IDAIF
	CBF	Δt	τ	ROIa1	ROIa2	ROIa3	Δ	Δ%

1	39.37	16.65	8.75	41.33	43.25	41.97	3.88	9.85
2	40.14	21.87	9.68	42.15	44.73	40.26	4.59	11.45
3	32.03	24.93	11.39	33.11	36.55	29.58	4.52	14.10
4	27.75	21.07	12.60	30.26	30.64	31.43	2.90	10.44
5	32.44	30.03	10.53	35.25	33.84	32.88	1.40	4.32
6	35.94	24.24	15.60	33.64	41.17	39.02	5.23	14.54
7	30.64	22.45	10.16	30.36	31.53	29.34	0.89	2.92
8	39.59	22.37	11.88	41.86	41.41	41.40	1.82	4.60
9	36.78	22.25	11.83	37.11	40.51	41.72	3.73	10.14
10	39.41	22.84	10.55	36.53	37.15	33.98	−2.26	−5.74
11a	25.14	21.41	9.28	17.74	17.28	20.32	−7.86	−31.25
12b	36.76	35.16	17.50	30.61	37.00	40.83	0.24	0.65
Mean	35.41	22.87	11.30	36.16	38.08	36.16	2.67	7.66
SDi	4.41	3.35	1.90	4.49	4.91	5.22	2.26	6.19

AIF, arterial input function; ASAIF, arterial sampling AIF; CBF, cerebral blood flow; ROI, region of interest; SDi, inter-subject standard deviation.

Not included in the mean and standard deviation analysis because of irregular low CBF, potentially because of excessive patient motion.

Not included in the mean and standard deviation analysis because of the unusually large dispersion constant.

Figure 5.

Example image-derived arterial input function (IDAIF), arterial sampling data after correction for delay and dispersion (ASAIF) and time-activity curves for whole brain (WBTAC). ASAIF, arterial sampling arterial input function; PET, positron emission tomography.

DISCUSSION

Reproducibility

The IDAIF technique generated highly reproducible whole-brain CBF measurements as assessed by CVw%. Presently obtained CVw% values were comparable to previously reported results for CBF measured with PET using conventional arterial sampling²³ as well as CBF measured with MRI using arterial spin labeling.²⁹ Whole-brain CBF measured on the two scanners (HR+ and PET/CT) was not only correlated but also virtually identical.

Image-Derived Arterial Input Function Versus Arterial Sampling Arterial Input Function

The presently described IDAIF method relies on empirical calibration of the effective scanner resolution to obtain results comparable to the conventional ASAIF technique. The presently tabulated results were obtained using an empirically derived scanner resolution of 8.3 mm FWHM, based on a single scan. Reasonably good agreement between IDAIF and the conventional technique (ASAIF) was observed, although IDAIF yielded systematically greater whole-brain CBF values (mean difference 2.67 mL/100 g per minute, or 7.66%). Recalibration based on all presently available group two data yielded an effective scanner resolution of 8.7 mm FWHM, eliminated systematic whole-brain CBF differences and increased ICC from 0.77 to 0.90. For the visual cortex gray matter, after recalibration, the ICC improved from 0.82 to 0.93, and the mean difference reduced to 0.74 mL/100 g per minute, or 1.31%.

Physical PET scanner resolution depends on gantry geometry,³⁰ the isotope being imaged³¹ and location in the field of view. Reported FWHM values for the HR+ scanner varied from 4.3 mm at the field of view center to 8.3 mm at the extreme periphery.³⁰ Another study reported that a 5-mm FWHM Gaussian filter was needed to smooth PET images acquired on an HR+ scanner to obtain 8 mm FWHM;²² this translates to a scanner resolution of ~6.2 mm FWHM. The difference between 6.2 mm FWHM and the presently obtained empirical value of 8.3 to 8.7 mm is likely attributable to errors in the arterial segmentation algorithm and the finite resolution of the time-of-flight MRA image: a segmentation procedure too generous in labeling voxels as the artery would cause overestimation of the recovery coefficient, r, in Equation (2). It should be noted that the spatial resolution of the image is also dependent on the reconstruction algorithm used, this may also be part of the reason the effective resolution based on our calibration was different from what was reported in the literature. Once the empirical calibration has been performed, the same image reconstruction method should be used for future studies to avoid biases. Regardless of the source of error, IDAIF is highly sensitive to the value used as the empirical PET scanner resolution: decreasing the FWHM by 0.5 mm results in overestimation of CBF by ~10%. One implication of these considerations is that consistent placement of subjects in the PET scanner will be important in applying IDAIF in practice. In the present work, if we had used the often-quoted value of 4.3 mm FWHM³⁰ in Equation (1), CBF would have been overestimated by ~90%. This effect probably accounts for the previously reported 104% overestimation of blood flow using an IDAIF technique.²⁵ It should be noted that the spatial resolution of PET scanner varies across the field of view; therefore, an effective resolution is determined by empirical calibration. This parameter is likely to be different for different choice of arterial ROIs, which is part of the reason for different CBF estimated as shown in Tables 1 and 2. However, different choice of aROI may be necessary, for example, in cases of subjects with partially occluded carotid arteries.

We used a constant water partition coefficient (0.90) within this study for all four regions used for the data fitting and parameter estimation. It is acknowledged that this parameter may differ from region to region, especially for the aROI. It is possible to include this as additional model parameter to be estimated, but it increases the instability of the procedure, therefore, it is not done in this study.

Uncertainties in Conventional Arterial Blood Sampling

Unless the AIF is sampled at the base of the brain, as in IDAIF, all AIF measurements must be corrected for delay and dispersion. The standard protocol at our institution³² involves deconvolution of the AIF measured at the radial artery by a hematocrit-dependent kernel derived by prior calibrations (Blood containing a tracer quantity of [¹⁵O]H₂O was placed in one-half of a chamber maintained at 37°C. Unlabeled blood was placed in the other half. The two halves were separated by a rubber septum. Blood was withdrawn at a constant rate of 5 mL/min through the same tubing used during PET scanning. The distal end of this tubing passed through a lead-shielded scintillation detector. The proximal end was connected to a needle. Calibration runs started with the needle in the cold chamber. To measure the convolution kernel of the tubing, the needle was passed through the septum, thereby generating a timed pulse of activity. Precise timing of the pulse was provided by a voltage level switched by the position of the needle.). This deconvolution is correct only if the blood sampling tubing and flow rate during PET scanning are identical to whatever has been calibrated. Moreover, the calibrated deconvolution accounts only for delay and dispersion within the tubing, not for the difference between arterial blood at the wrist versus that supplying the brain. These problems with ASAIF may partly account for the anomalous results obtained with data set #12. To avoid these problems, we adopted a modeling approach^25–27 to correct for ASAIF delay and dispersion. It is observed that the estimated delay and dispersion parameters are larger than what has been shown in other studies.²⁶ This is likely because of the difference in the automated blood sampling device.

Dependence of Measurement on the Evaluation Interval

Evaluation of Equation (3) was restricted to 40-second window, as in our standard quantitative [¹⁵O]H₂O auto-radiographic protocol.^3,6,32 It has been observed that using a longer time window leads to underestimation of blood flow,^3,33 which result has been attributed to the complicated kinetics of water uptake within the brain.³⁴ Other groups have used longer evaluation times.^35–37 Here, we note that estimating CBF with IDAIF over extended dynamic data sets did not markedly reduce estimated CBF, whereas with ASAIF, it did. As can be seen in Figure 5, ASAIF was consistently above the IDAIF after the main peak; overestimation of the AIF results in lower estimated CBF. We hypothesize that the delay and dispersion correction model for the ASAIF technique is oversimplified and may cause overestimation of the AIF at a later time point, while the IDAIF more correctly models the compartmental kinetics of H₂O uptake within the brain. In fact, using a longer time window may be beneficial since more data are available. It should be noted that the kinetic model used in this study as well as many [¹⁵O]-water-based CBF studies assumes a complete extraction, which is not completely true especially under high flow conditions.^38,39 This simplification will lead to underestimation of the CBF, which is another source of error in [¹⁵O]-water-based CBF technique.

Sensitivity to Image Registration Error

Image registration error introduces noise in estimating CBF using the IDAIF approach. Deliberately introducing registration errors showed that a translational error of 1 mm or a rotational error of 1° leads to underestimation of the AIF and ~10% overestimation of CBF. However, transform consistency-based evaluation of our registration error⁴⁰ indicates that we routinely achieve errors under 0.5 mm translation and 0.5° rotation. This implies that computational registration error does not present a major problem in our AIF estimation technique. Subject motion during PET scanning is another factor that can affect CBF estimation. The IDAIF is more sensitive to subject motion than ASAIF because IDAIF estimates the AIF by differentiating the whole-brain TAC, which becomes noisy consequent to subject motion. In fact, subject motion could well account for the anomalously low global CBF in data set #11. Alternative approaches may reduce sensitivity to patient motion. For example, the AIF may be estimated by solving a set of linear equations in a manner similar to the hybrid-single-input-dual-output (HSIDO) technique we proposed for animal studies.¹⁴ In addition, one may also correct for motion using a registration technique.⁴¹ Image registration errors introduced by subject motion may be minimized by use of a PET/MR hybrid scanner.⁴² We currently are evaluating the IDAIF method on a Biograph mMR scanner (Siemens Medical Solutions USA, Inc.).

We adopted high-resolution MRA to facilitate AIF estimation based on the fact that MRI is routinely performed in many neurologic imaging studies, clinical as well as experimental. Adding a 9-minute MRA scan is not unduly burdensome to patients nor would it dramatically increase MRI costs. The same IDAIF approach can be applied using CT angiography. While this has the potential benefit of obtaining an arterial image using a single PET/CT scanner, it would substantially increase the radiation dose to obtain the high-quality image required by IDAIF. It should be noted that we have shown comparable CBF measurements using a PET/CT scanner, which operates in 3D mode, and a (PET only) HR+ scanner operating in 2D mode. Thus, the increased scattering associated with 3D PET imaging does not appear to affect CBF estimation using IDAIF.

Summary

In summary, an IDAIF estimation technique using coregistered MRA images was developed and validated in the context of quantitative [¹⁵O]H₂O CBF imaging. The IDAIF technique offers several advantages in comparison with conventional arterial sampling. First, it does not require invasive arterial catheterization, which is burdensome to participants and can fail in >10% of cases for various reasons.⁴³ Second, the arterial blood samples are commonly taken from an artery that does not supply the brain, which introduces measurement noise and bias. Moreover, IDAIF is suitable for use during combined PET-MRI, whereas special instrumentation would be required to measure the ADAIF in the presence of a large, static magnetic field. However, IDAIF is sensitive to several factors, viz., the modeled effective scanner resolution and image registration error including that caused by subject motion. Nevertheless, we show that the IDAIF method provides reproducible CBF estimation without arterial blood sampling. Good agreement in CBF measured using IDAIF versus conventional arterial sampling also was observed. Further work is ongoing to validate the IDAIF technique for use with other tracers such as [¹⁵O]O₂ and [¹⁵O]CO.

DISCLOSURE/CONFLICT OF INTEREST

The authors declare no conflict of interest.

References

Wintermark

Sesay

Barbier

Borbely

Dillon

Eastwood

Comparative overview of brain perfusion imaging techniques. Stroke 2005; 36: e83–e99.

Mintun

Raichle

Martin

Herscovitch

. Brain oxygen utilization measured with O-15 radiotracers and positron emission tomography. J Nucl Med 1984; 25: 177–187.

Raichle

Martin

Herscovitch

Mintun

Markham

. Brain blood flow measured with intravenous H2(15)O. II. Implementation and validation. J Nucl Med 1983; 24: 790–798.

Donahue

Jones

Pekar

Van Zijl

. An account of the discrepancy between MRI and PET cerebral blood flow measures. A high-field MRI investigation. NMR Biomed 2006; 19: 1043–1054.

Martin

Powers

Raichle

. Cerebral blood volume measured with inhaled C15O and positron emission tomography. J Cereb Blood Flow Metab 1987; 7: 421–426.

Videen

Perlmutter

Herscovitch

Raichle

. Brain blood volume, flow, and oxygen utilization measured with 15O radiotracers and positron emission tomography: revised metabolic computations. J Cereb Blood Flow Metab 1987; 7: 513–516.

Eberl

Anayat

Fulton

Hooper

Fulham

. Evaluation of two population-based input functions for quantitative neurological FDG PET studies. Eur J Nucl Med 1997; 24: 299–304.

Takikawa

Dhawan

Spetsieris

Robeson

Chaly

Dahl

Noninvasive quantitative fluorodeoxyglucose PET studies with an estimated input function derived from a population-based arterial blood curve. Radiology 1993; 188: 131–136.

Croteau

Lavallée

Labbe

Hubert

Pifferi

Rousseau

Image-derived input function in dynamic human PET/CT: methodology and validation with 11C-acetate and 18F-fluorothioheptadecanoic acid in muscle and 18F-fluorodeoxyglucose in brain. Eur J Nucl Med Mol Imaging 2010.

10.

Mourik

JEM

Lubberink

Klumpers

UMH

Comans

Lammertsma

Boellaard

. Partial volume corrected image derived input functions for dynamic PET brain studies: Methodology and validation for [11C]flumazenil. Neuroimage 2008; 39: 1041–1050.

11.

Sitek

Gullberg

Huesman

. Correction for ambiguous solutions in factor analysis using a penalized least squares objective. IEEE Trans Med Imaging 2002; 21: 216–225.

12.

Houston

. The effect of apex-finding errors on factor images obtained from factor analysis and oblique transformation. Phys Med Biol 1984; 29: 1109–1116.

13.

Welch

Shoghi

. The application of maximum likelihood factor analysis (MLFA) with uniqueness constraints on dynamic cardiac microPET data. Phys Med Biol 2007; 52: 2313–2334.

14.

Shoghi

. Single-input-dual-output modeling of image-based input function estimation. Mol Imaging Biol 2010; 12: 286–294.

15.

Fang

Muzic Jr.

RF. Spillover and partial-volume correction for image-derived input functions for small-animal 18F-FDG PET studies. J Nucl Med 2008; 49: 606–614.

16.

Feng

Wong

Siu

. A technique for extracting physiological parameters and the required input function simultaneously from PET image measurements: theory and simulation study. IEEE Trans Inf Technol Biomed 1997; 1: 243–254.

17.

Zanotti-Fregonara

Fadaili

Maroy

Comtat

Souloumiac

Jan

Comparison of eight methods for the estimation of the image-derived input function in dynamic [lsqb]18F[rsqb]-FDG PET human brain studies. J Cereb Blood Flow Metab 2009; 29: 1825–1835.

18.

Quarles

Mintun

Larson

Markham

MacLeod

Raichle

. Measurement of regional cerebral blood flow with positron emission tomography: a comparison of [15O]water to [11C]butanol with distributed-parameter and compartmental models. J Cereb Blood Flow Metab 1993; 13: 733–747.

19.

Rowland

Garbow

Laforest

Snyder

. Registration of [18F]FDG microPET and small-animal MRI. Nucl Med Biol 2005; 32: 567–572.

20.

Hajnal

Saeed

Soar

Oatridge

Young

Bydder

. A registration and interpolation procedure for subvoxel matching of serially acquired MR images. J Comput Assist Tomogr 1995; 19: 289–296.

21.

Wilson

Noble

. An adaptive segmentation algorithm for time-of-flight MRA data. IEEE Trans Med Imaging 1999; 18: 938–945.

22.

Joshi

Koeppe

Fessler

. Reducing between scanner differences in multicenter PET studies. Neuroimage 2009; 46: 154–159.

23.

Law

Iida

Holm

Nour

Rostrup

Svarer

Quantitation of regional cerebral blood flow corrected for partial volume effect using O-15 water and PET[colon] II. Normal values and gray matter blood flow response to visual activation. J Cereb Blood Flow Metab 2000; 20: 1252–1263.

24.

Herscovitch

Raichle

. What is the correct value for the brain–blood partition coefficient for water? J Cereb Blood Flow Metab 1985; 5: 65–69.

25.

Ludemann

Sreenivasa

Michel

Rosner

Plotkin

Felix

Corrections of arterial input function for dynamic H215O PET to assess perfusion of pelvic tumours: arterial blood sampling versus image extraction. Phys Med Biol 2006; 51: 2883–2900.

26.

Meyer

. Simultaneous correction for tracer arrival delay and dispersion in CBF measurements by the H215O autoradiographic method and dynamic PET. J Nucl Med 1989; 30: 1069–1078.

27.

Iida

Kanno

Miura

Murakami

Takahashi

Uemura

. Error analysis of a quantitative cerebral blood flow measurement using H2(15)O autoradiography and positron emission tomography, with respect to the dispersion of the input function. J Cereb Blood Flow Metab 1986; 6: 536–545.

28.

Shrout

Fleiss

. Intraclass correlations: uses in assessing rater reliability. Psychol Bull 1979; 86: 420–428.

29.

Wang

Saykin

Pfeuffer

Lin

Mosier

Shen

Regional reproducibility of pulsed arterial spin labeling perfusion imaging at 3T. Neuroimage 2011; 54: 1188–1195.

30.

Brix

Zaers

Adam

Bellemann

Ostertag

Trojan

Performance evaluation of a whole-body PET scanner using the NEMA protocol. National Electrical Manufacturers Association. J Nucl Med 1997; 38: 1614–1623.

31.

Levin

Hoffman

. Calculation of positron range and its effect on the fundamental limit of positron emission tomography system spatial resolution. Phys Med Biol 1999; 44: 781–799.

32.

Powers

Videen

Markham

Walter

Perlmutter

. Metabolic control of resting hemispheric cerebral blood flow is oxidative, not glycolytic. J Cereb Blood Flow Metab 2011; 31: 1223–1228.

33.

Eklof

Lassen

Nilsson

Norberg

Siesjo

Torlof

. Regional cerebral blood flow in the rat measured by the tissue sampling technique; a critical evaluation using four indicators C14-antipyrine, C14-ethanol H3-water and xenon. Acta Physiol Scand 1974; 91: 1–10.

34.

Larson

Markham

Raichle

. Tracer-kinetic models for measuring cerebral blood flow using externally detected radiotracers. J Cereb Blood Flow Metab 1987; 7: 443–463.

35.

Ito

Kanno

Kato

Sasaki

Ishii

Ouchi

Database of normal human cerebral blood flow, cerebral blood volume, cerebral oxygen extraction fraction and cerebral metabolic rate of oxygen measured by positron emission tomography with 15O-labelled carbon dioxide or water, carbon monoxide and oxygen: a multicentre study in Japan. Eur J Nucl Med Mol Imaging 2004; 31: 635–643.

36.

Shidahara

Watabe

Kim

Kudomi

Ito

Iida

. Optimal scan time of oxygen-15-labeled gas inhalation autoradiographic method for measurement of cerebral oxygen extraction fraction and cerebral oxygen metabolic rate. Ann Nucl Med 2008; 22: 667–675.

37.

Ohta

Meyer

Fujita

Reutens

Evans

Gjedde

. Cerebral [15O]Water clearance in humans determined by PET: I. Theory and normal values. J Cereb Blood Flow Metab 1996; 16: 765–780.

38.

Silva

Zhang

Williams

Koretsky

. Estimation of water extraction fractions in rat brain using magnetic resonance measurement of perfusion with arterial spin labeling. Magn Reson Med 1997; 37: 58–68.

39.

Takagi

Ehara

Finn

. Water extraction fraction and permeability-surface product after intravenous injection in rats. Stroke 1987; 18: 177–183.

40.

Eisenstein

Koller

Piccirillo

Kim

Antenor-Dorsey

Videen

Characterization of extrastriatal D2 in vivo specific binding of [(18) F] (N-methyl)benperidol using PET. Synapse 2012; 66: 770–780.

41.

. Inter-frame motion correction for small animal PET imaging. Biomedical Engineering and Informatics (BMEI). 2011 4th International Conference 2011; 1: 338–342.

42.

Delso

Furst

Jakoby

Ladebeck

Ganter

Nekolla

Performance Measurements of the Siemens mMR Integrated Whole-Body PET/MR Scanner. J Nucl Med 2011; 52: 1914–1922.

43.

Derdeyn

Videen

Simmons

Yundt

Fritsch

Grubb

Count-based PET method for predicting ischemic stroke in patients with symptomatic carotid arterial occlusion. Radiology 1999; 212: 499–506.

Noninvasive Estimation of the Arterial Input Function in Positron Emission Tomography Imaging of Cerebral Blood Flow

Abstract

INTRODUCTION

MATERIALS AND METHODS

Human Subjects

Image Processing

Image-Derived Arterial Input Function Method

Cerebral Blood Flow Calculation with Arterial Sampling (Arterial Sampling Arterial Input Function)

Statistical Analysis

RESULTS

Study Group 1: Within-Subject Reproducibility and Comparison of Two Scanners

Study Group 2: Image-Derived Arterial Input Function Versus Conventional Arterial Sampling

DISCUSSION

Reproducibility

Image-Derived Arterial Input Function Versus Arterial Sampling Arterial Input Function

Uncertainties in Conventional Arterial Blood Sampling

Dependence of Measurement on the Evaluation Interval

Sensitivity to Image Registration Error

Summary

DISCLOSURE/CONFLICT OF INTEREST

References