Cerebral blood flow and glucose metabolism in healthy volunteers measured using a high-resolution PET scanner

Positron emission tomography (PET) measurements using ¹⁸ F]-2-fluoro-2-deoxy-d-glucose (¹⁸ F]FDG) allow for quantitative determination of the cerebral metabolic rate of glucose utilization (CMR_glu) [1, 2]. Although, initially, rate constants were measured [3–5] using the compartment model developed by Sokoloff and co-workers [6], thereafter, CMR_glu measurements were increasingly based on the assumption of fixed rate constants in combination with static scans [7]. As a result, no recent estimates of the rate constants obtained using a state-of-the-art high-resolution scanner are available for human studies (estimates have only been reported for mice; see [8]).

Quantification of regional cerebral blood flow (CBF) based on ¹⁵O]H₂O and PET also started around the same time period as metabolic assessments, i.e., around 1980. Due to hardware limitations, scanners were too slow for fast dynamic scans, leading to steady-state [9] and autoradiographic protocols [10]. Dynamic scan protocols were not introduced until the late 1980s [11, 12], but to date, relatively few methodological papers on fully dynamic implementations have been published. In a review by Silverman and Phelps, it was noted that CBF measurements showed more variability than CMR_glu measurements with standard deviations within investigations of around 10% to 20% and up to 100% between investigations [13].

The high-resolution research tomograph (HRRT; CTI/Siemens, Knoxville, TN, USA) is a dedicated brain scanner that combines high spatial resolution (approximately 3 mm) with good sensitivity [14]. Previously, it has been shown that the increased spatial resolution allows for measurements of CMR_glu with reduced underestimation due to partial volume effects [15, 16]. These glucose measurements, however, were based on a static FDG protocol together with population average-based values for the rate constants. In a recent study, HRRT image quality was shown to be good enough to allow for quantitatively correct CBF measurements [17].

The use of an arterial line (blood sampler derived input function (BSIF) to acquire an input function is the gold standard for dynamic data analysis of cerebral studies, but it is an invasive procedure. Therefore, use of an image-derived input function (IDIF) can be an interesting alternative [18–20], but its utility needs to be validated for each tracer, each scanner, and each acquisition and data analysis protocol separately.

The main purpose of the present study was to derive human CBF and CMR_glu values as measured using a current state-of-the-art high-resolution scanner. Data were analyzed by full kinetic analysis using nonlinear regression (NLR) of regional time-activity curve (TAC) data with a BSIF. A second objective was to assess the accuracy of high-resolution parametric images of CBF and CMR_glu. In addition, CBF and K _i values for automatically delineated regions were compared with manually drawn regions of interest and literature values. Based on earlier work on IDIFs [18, 20, 21] and given the high resolution of the HRRT, an additional objective was to assess whether a carotid artery-based IDIF could be used as a noninvasive alternative for arterial sampling in the case of both CBF and CMR_glu measurements, thereby increasing clinical applicability of this methodology in humans.

To date, no full kinetic analysis of [¹⁸ F]FDG data based on arterial sampling and a dynamic HRRT scanning protocol in humans has been reported. In the present study, average CMR_glu values of 0.33 and 0.11 μmol/cm³ per minute were obtained for manually drawn gray and white matter regions, respectively. Automatically delineated regions yielded values of 0.29 and 0.19 μmol/cm³ per minute for gray matter and white matter, respectively, averaged over nine patients with a CoV of less than 15%. However, especially the white matter estimate based on the latter method was contaminated with some gray matter spill in.

Previously, Heiss and co-workers have acquired CMR_glu data on an HRRT scanner [15]. Using fixed rate constants from the literature, they found similar gray and white matter CMR_glu values with a GM/WM ratio of 2.7, which is only slightly lower than the present ratio for manually drawn regions. Therefore, it seems that a static scanning protocol, together with fixed rate constants, is a valid approximation of full kinetic modeling in the case of healthy volunteers. Nevertheless, it should be noted that the assumption of fixed rate constants may not be valid in certain clinically relevant patient populations, such as those having diabetes [31], characterized by a.o. hyperglycemia and hyperinsulinemia which both affect the blood vessel wall and the permeability surface area product [32], and decreased glucose metabolism [33]. Furthermore, normal fixed rate constants are not automatically applicable to flow-limited states [34].

The good correlation between Patlak and NLR results (slope 0.96, r² 0.98) implies that this linearization is a valid approach, as shown previously [7]. The underestimation of 4% is likely due to the blood volume (5%) that is taken into account in the NLR analysis, but not in the Patlak linearization. Parametric K _i images without smoothing showed good correlation with regional K _i values (r² of 0.99) with a slope of 1.04, probably induced by differences in noise and tissue heterogeneity present in regional- versus voxel-based TACs. Smoothing the dynamic ¹⁸ F]FDG images resulted in a lower slope and poorer correlation, and this is therefore not recommended.

Average CBF values of 0.43 and 0.13 mL/cm³ per minute were obtained for manually drawn gray and white matter regions, respectively. These values are in line with the recent data acquired on an HRRT by Walker and co-workers [17], who found values of 0.44 and 0.15 mL/cm³ per minute (Table 3). The good correlation between BFM and NLR results implies that this linearization is a valid approach.

We found an FDG extraction fraction of 18%, which is in line with earlier reported values of approximately 20% [35], determined using the double-indicator method [36]. To our knowledge, this has not been derived from a combined CBF and CMR_glu (K ₁) measurement before.

The use of a noninvasive input function, derived from the carotid arteries, allowed for quantitatively correct estimates of regional CMR_glu when Patlak linearization was used. In the case of NLR, however, the more stringent requirements placed on the input function (i.e., in addition to the area under the curve, the detailed shape of the peak is needed) resulted in good CMR_glu estimates for only seven out of nine subjects and incorrect (with an average slope of 0.76) estimates in the other two. Nevertheless, clinical dynamic [¹⁸ F]FDG studies without the need for an arterial line can be performed (on a voxel-by-voxel basis) by analyzing data using the Patlak linearization.

Unfortunately, the IDIF approach was not applicable to the analysis of CBF scans. Scaling to the manual samples yielded a similar factor as for CMR_glu scans, but an underestimation of the peak was observed (by a factor of 3). Although ¹⁵O has a higher positron energy (leading to an effective spatial resolution of 3.4 in the case of ¹⁵O, if it were 3.0 for ¹⁸ F), this is not likely to be the main reason for the underestimation of the peak. Possible explanations will probably include the implementation of scatter correction as well as the performance of the reconstruction algorithm for highly localized activity distributions, as prevailing in the blood just after the [¹⁵O]H₂O bolus. Therefore, optimization of the frame definition and the use of a point spread function reconstruction may lead to a better match between BSIF and IDIF.

It is of importance to note that, although the images used were smoothed 6 mm at FWHM, the IDIF applicability cannot be automatically used for data generated at other PET scanners. This still needs a validation for every single tracer, every single scanner, and each acquisition and data analysis protocol separately. The success of the use of an IDIF depends on the intrinsic spatial resolution of the scanner, which is often lower than that of the scanner used in our study, and on the iterative reconstruction algorithm used (e.g., whether or not priors are included and whether the partial volume effect is implemented in the reconstruction algorithm), determining the signal-to-noise ratio as well as the minimum frame duration that can be applied. Furthermore, the duration of tracer injection needs to be optimized since the number of frames that can be acquired during the bolus should not be too low. The exact effect of these different options cannot be predicted and needs to be tested in the way as was described in this paper.

NLR-based regional CMR_glu and CBF values can be obtained using a high-resolution PET brain scanner in healthy humans. Good correspondence between the nonlinear and linear models was observed for the regional data, which formed the basis for the calculation of accurate parametric images. Although no test-retest study was performed, average inter-subject variability of approximately 15% was observed, which implies a good reproducibility. For studies that would only require measurements of CMR_glu, regional Patlak-based K _i values can be derived without arterial sampling by using a carotid-based image-derived input function.