Evaluation of Arterial Spin Labeling MRI—Comparison with 15O-Water PET on an Integrated PET/MR Scanner

Cerebral blood flow (CBF) measurements are of high clinical value and can be acquired non-invasively with no radiation exposure using pseudo-continuous arterial spin labeling (ASL). The aim of this study was to evaluate accordance in resting state CBF between ASL (CBFASL) and 15O-water positron emission tomography (PET) (CBFPET) acquired simultaneously on an integrated 3T PET/MR system. The data comprised ASL and dynamic 15O-water PET data with arterial blood sampling of eighteen subjects (eight patients with focal epilepsy and ten healthy controls, age 21 to 61 years). 15O-water PET parametric CBF images were generated using a basis function implementation of the single tissue compartment model. Cortical and subcortical regions were automatically segmented using Freesurfer. Average CBFASL and CBFPET in grey matter were 60 ± 20 and 75 ± 22 mL/100 g/min respectively, with a relatively high correlation (r = 0.78, p < 0.001). Bland-Altman analysis revealed poor agreement (bias = −15 mL/100 g/min, lower and upper limits of agreements = −16 and 45 mL/100 g/min, respectively) with a negative relationship. Accounting for the negative relationship, the width of the limits of agreement could be narrowed from 61 mL/100 g/min to 35 mL/100 g/min using regression-based limits of agreements. Although a high correlation between CBFASL and CBFPET was found, the agreement in absolute CBF values was not sufficient for ASL to be used interchangeably with 15O-water PET.


Introduction
Cerebral blood flow (CBF) measurements are of high clinical value for various brain disorders such as cerebrovascular disorders, brain tumors, and neurodegenerative diseases [1][2][3]. Current reference standard for CBF measurements is positron emission tomography (PET) with 15 O-water [3]. Although 15 O-water PET has proven its usefulness in physiological experiments and clinical assessments [4], it is often considered costly and implementations are limited due to the requirement for an on-site cyclotron and an arterial line for blood sampling. Arterial spin labeling (ASL) is a non-invasive magnetic resonance imaging (MRI)-based CBF measurement technique using the patient's own water molecules in blood as a freely diffusible tracer. ASL is not a novel technique; the basic principle was already introduced in the early 1990s [5][6][7]. A consensus statement was published in 2015 [8] recommending pseudo-continuous ASL as a labeling strategy with 3D segmented read-out applying background suppression [8].
There are a number of methodological considerations when comparing 15 O-water PET and ASL acquired CBF. First, comparison studies should be preferably based on si-multaneously acquired 15 O-water PET and ASL CBF measurements using an integrated PET/MR system since the physiological cerebral perfusion state is time dependent [1,[9][10][11]. Second, the lack of CT data for attenuation correction of PET images on an integrated PET/MR system needs another approach. Although MRI-based attenuation correction methods are still under investigation, zero-echo time (ZTE)-based attenuation correction has recently been demonstrated to be an adequate method for brain PET/MR applications [12][13][14]. Third, 15 O-water PET is considered the gold standard for measuring CBF when using a kinetic modeling approach with an arterial input function (AIF) based on arterial sampling [1,15,16]. Although several studies compared CBF measurements with 15 O-water PET and ASL, none of them fulfilled all three requirements. From those studies conducted on a PET/MR system, the MRI-based attenuation correction methods employed have generally demonstrated inadequate performance compared to ZTE-based attenuation correction. Further, arguing its invasiveness and discomfort, arterial blood sampling was often omitted [9,17]. Moreover, there were obvious differences in study participants and the ASL method used.
The aim of this study was to evaluate accordance in resting state CBF values based on simultaneously acquired ASL (CBF ASL ) and 15 O-water PET (CBF PET ) on an integrated PET/MR system and using ZTE-based attenuation correction and arterial sampling.

Scope and Subjects
In this methodological study, conducted between December 2015 and May 2018, a comparative analysis was performed on, in total, eighteen participants-eight patients with focal epilepsy (5 females, 3 males) with a mean (standard deviation, SD) age of 39 (13) and ten healthy controls (5 females, 5 males) with a mean (SD) age of 40 (12). The patients were given lamotrigine (4 patients) or carbamazepine (6 patients). Two patients were given carbamazepine in combination with clonazepam or valproic acid. No patient had a history of taking levetiracetam. Furthermore, none of the enrolled participants had any intellectual disability. The healthy controls were matched on age and sex with the patients. Recruitment was done by advertisement and no participant had any relations to the hospital or the faculty staff. Each participant was its own control, and thus potential regional differences in CBF due to factors like age, gender and groups of participants were outside the scope of the study. Epilepsy patients were scanned in interictal state.
The study was done in accordance with the declaration of Helsinki. Approvals were obtained by the Regional Board of Medical Ethics in Uppsala (DNR 2015/187) and the Radiation Ethics Committee at Uppsala University Hospital. After a complete description of the study, and prior to inclusion, all participants signed an informed consent form.

Data Acquisition
All examinations were performed on an integrated PET/MR (SIGNA, GE Healthcare, Waukesha, WI, USA) which combines a 3T MRI with a time-of-flight capable silicone photomultiplier-based PET scanner [18]. All subjects were scanned in supine position using an eight-channel head coil (MR Instruments Inc., Minneapolis, MN, USA). A 10-min dynamic PET scan was started after automatic bolus injection (1 mL/s during 5 s) of 5 MBq/kg 15 O-water (max 500 MBq), followed by flushing with 35 mL saline at 2 mL/s. Continuous blood sampling (3 mL/min) was conducted from a radial artery, generally in the non-dominant arm, and blood radioactivity was measured using a Twilite Two blood detector (Swisstrace, Zurich). The blood detector was positioned on the scanner bed as close as possible to the subjects' wrists to minimise dispersion. During PET scanning, a 3D pseudo-continuous ASL with background suppressed fast spin echo spiral read-out using a PLD of 2025 ms and label duration of 1800 ms was acquired. In addition, the protocol included a high-resolution 3D-T1-weighted (T1w) image and a 3D-T2-weighted fluid attenuated inversion recovery (T2w-FLAIR) as anatomical references and a ZTE image for attenuation correction of PET data. The full set of acquisition parameters for MRI scans is presented in Supplementary Table S1.

Image Reconstruction and Generation of Parametric CBF Images
The 15 O-water PET images were reconstructed using time-of-flight ordered subset expectation maximization (4 iterations, 28 subsets), with ZTE-based attenuation correction [13,14] and a 5 mm Gaussian post-filter into 22 frames of increasing durations (1 × 10 s, 8 × 5 s, 4 × 10 s, 2 × 15 s, 3 × 20 s, 2 × 30 s, 2 × 60 s) into a 128 × 128 × 89 matrix with 2.34 × 2.34 × 2.81 mm 3 voxels. The last four minutes of the acquisition were not included due to limited information and increasingly noisy blood data. All subject-specific AIFs were corrected for delay and dispersion [19] and 15 O-water PET derived parametric CBF images (CBF PET ) were produced using a basis function implementation of the standard single-tissue compartment model including a fitted blood volume parameter [20]. ASL derived parametric CBF images (CBF ASL ) were generated according the single compartment model defined by Buxton et al., [21] and recommended by Alsop et al., [8] including a correction term for full proton density reference [22].

Post-Processing
T2w-FLAIR, CBF ASL and CBF PET -images were co-registered to each subject's corresponding T1w images. Grey matter (GM) tissue probability maps were segmented based on T1w images and co-registered T2w-FLAIR images. GM maps were defined with a tissue probability fraction above 75%. White matter (WM) maps were disregarded because of the general limitations of ASL in WM [1,8]. All processing steps, as described above, were performed using the SPM12 toolbox (Wellcome Trust Centre for Neuroimaging, London, UK).

Comparative Analysis
A descriptive analysis was made to compare CBF ASL and CBF PET for both all selected brain regions as well as for clusters of cortical and subcortical regions and whole-brain GM. Agreement between quantitative CBF ASL and CBF PET values was first studied using correlation analysis including Pearson's product moment correlation coefficient and orthogonal regression. The correlation and regression measures are reported with corresponding 95% confidence interval (CI). Thereafter, Bland-Altman analyses were performed to examine the relationship between the average of CBF ASL and CBF PET and the difference between CBF ASL and CBF PET . In addition, bias, expressed as average difference between both methods, was estimated with 95% lower and upper limits of agreement (LoA L and LoA U , respectively). Potential relationships between the difference and the average of CBF PET and CBF ASL were identified using linear regression [24], and regression-based LoAs (RLoA L and RLOA U ) were calculated to account for any potential relationships found [25]. A calculation example for GM is provided in Supplementary Document S1. All statistical tests are two-sided using GraphPad Prism 8 (GraphPad Software, La Jolla, CA, USA).

Descriptive Analysis
Average parametric CBF ASL and CBF PET images and the differences between both methods are shown in Figure 1. This figure illustrates that CBF ASL resulted in lower values than CBF PET , which was consistent throughout the whole GM. Quantitative CBF values are presented for all regions and both methods in Table 1. In GM, average CBF was 75 ± 22 and 60 ± 10 mL/100 g/min for CBF PET and CBF ASL , respectively. A larger average difference between CBF ASL and CBF PET was found in subcortical regions compared to cortical regions, especially in caudate, putamen, pallidum, and thalamus. Further, it can be noticed that the variability across subjects, given as SD, was substantially greater for CBF PET compared to CBF ASL .

Descriptive Analysis
Average parametric CBFASL and CBFPET images and the differences between both methods are shown in Figure 1. This figure illustrates that CBFASL resulted in lower values than CBFPET, which was consistent throughout the whole GM. Quantitative CBF values are presented for all regions and both methods in Table 1. In GM, average CBF was 75 ± 22 and 60 ± 10 mL/100 g/min for CBFPET and CBFASL, respectively. A larger average difference between CBFASL and CBFPET was found in subcortical regions compared to cortical regions, especially in caudate, putamen, pallidum, and thalamus. Further, it can be noticed that the variability across subjects, given as SD, was substantially greater for CBFPET compared to CBFASL.

Correlation and Regression
Correlations between CBFASL and CBFPET are shown for GM as well as clusters of cortical-and subcortical regions in Figure 2. Correlations between CBFASL and CBFPET are given for all regions in Table 1. Similar and positive correlations between CBFASL and

Correlation and Regression
Correlations between CBF ASL and CBF PET are shown for GM as well as clusters of cortical-and subcortical regions in Figure 2. Correlations between CBF ASL and CBF PET are given for all regions in Table 1

Analysis of Agreement between CBFASL and CBFPET
Bland-Altman plots displayed a negative, nearly linear relationship for GM ( Figure  3a) as well as cortical-and subcortical regions (Figure 3b,c, respectively). Bias was −15, −13 and −20 mL/100 g/min for GM, cortical-and subcortical regions, respectively. Quantitative results of the Bland-Altman analysis are given for individual regions in Table 2. The differences between the considered cortical regions were relatively small. In contrast, subcortical regions showed a large variation. The highest bias was found in putamen, pallidum, and thalamus (−34 to −26 mL/100 g/min) and lowest bias in amygdala and hippocampus (−10 to −6 mL/100 g/min).

Regression-Based Limits of Agreements
A negative relationship of the difference on the average of CBFPET and CBFASL was found and RLoAs were calculated for all regions. The bias remains unchanged using RLoAs. The width between upper and lower LoAs changed from 61 mL/100 g/min to 35 mL/100 g/min in GM using RLoAs. In cortical-and subcortical regions, the width changed from 62 mL/100 g/min to 40 mL/100 g/min and from 71 mL/100 g/min to 46 mL/100 g/min, respectively (compare

Analysis of Agreement between CBF ASL and CBF PET
Bland-Altman plots displayed a negative, nearly linear relationship for GM (Figure 3a) as well as cortical-and subcortical regions (Figure 3b,c, respectively). Bias was −15, −13 and −20 mL/100 g/min for GM, cortical-and subcortical regions, respectively. Quantitative results of the Bland-Altman analysis are given for individual regions in Table 2. The differences between the considered cortical regions were relatively small. In contrast, subcortical regions showed a large variation. The highest bias was found in putamen, pallidum, and thalamus (−34 to −26 mL/100 g/min) and lowest bias in amygdala and hippocampus (−10 to −6 mL/100 g/min). for all regions when using RLoAs instead of ordinary LoAs. Quantitative results from the RLoA method are given for individual regions in Table 3.

Regression-Based Limits of Agreements
A negative relationship of the difference on the average of CBF PET and CBF ASL was found and RLoAs were calculated for all regions. The bias remains unchanged using RLoAs. The width between upper and lower LoAs changed from 61 mL/100 g/min to 35 mL/100 g/min in GM using RLoAs. In cortical-and subcortical regions, the width changed from 62 mL/100 g/min to 40 mL/100 g/min and from 71 mL/100 g/min to 46 mL/100 g/min, respectively (compare Figure 3a-c with Figure 3c-f). A consistent narrowing was found for all regions when using RLoAs instead of ordinary LoAs. Quantitative results from the RLoA method are given for individual regions in Table 3. Table 3. Slope and intercept of the upper and lower regression-based limits of agreement (RLOA U and RLOA L , respectively) with slope and intercept from the regression analysis of the difference between CBF ASL -CBF PET on average of CBF ASL + CBF PET . A calculation example is given in Supplementary Document S1.

Region
Linear

Discussion
This study evaluated the agreement between CBF ASL and CBF PET derived from parametric images allowing a quantitative comparison between both methods. We found a relatively high correlation between CBF ASL and CBF PET in GM in comparison to previously published work. However, the agreement between CBF ASL and CBF PET was poor. We observed a negative proportional bias between the difference and average of CBF PET and CBF ASL in all regions. This is also apparent in the orthogonal regression, where the slope is less than 1 for all regions, i.e., the difference will increase as the average of CBF PET and CBF ASL increases. There is also an apparent underestimation of CBF ASL in subcortical regions, which may be caused by shorter than assumed T1 relaxation and earlier arrival of labeled blood compared to cortical regions [9,22].
Previous studies comparing ASL and 15 O-water PET have reported correlation coefficients in GM ranging from 0.26 to 0.81 [4,10,11,[26][27][28][29][30][31][32][33] (see Figure 4). In our study, we found correlations between CBF ASL and CBF PET varying between 0.42 in pallidum and 0.83 in frontal cortex. Two of the previous studies performed arterial blood sampling and used an integrated PET/MR. Zhang et al. [11]. reported a correlation of 0.80 in GM and 0.61 to 0.87 in cortical-and subcortical regions comparing simultaneously acquired CBF measurements. Although the reported correlation coefficients agree with ours, Zhang et al. [11]. found generally higher values for CBF ASL than for CBF PET , which is contrary to our results. However, Zhang et al. [11], used template-based attenuation correction, which might underestimate PET tracer uptake [14,34]. Puig et al. [4]. compared simultaneously acquired CBF measurements during rest in healthy subjects and found a correlation of 0.32. Correlations in cortical-and subcortical regions were reported using combined data from rest and altered perfusion states, so any comparisons to our study are hard to make. Both Zhang et al. [11] and Puig et al. [4] also performed a Bland-Altman analysis and reported a bias (LoA L and LoA U ) of 15 (−5 and 25, width 30) and 0 (−15 and 15, width 30) mL/100 g/min in GM [4,11], respectively, compared to −15 (−45 and 16, width 61) mL/100 g/min in the present study. No regression analysis was performed in any of the two studies. Furthermore, Bland-Altman analysis of cortical and subcortical regions was omitted or included several altered perfusion states in the two above mentioned studies, so no further comparison to our results is possible except for whole GM. previously published work. However, the agreement between CBFASL and CBFPET was poor. We observed a negative proportional bias between the difference and average of CBFPET and CBFASL in all regions. This is also apparent in the orthogonal regression, where the slope is less than 1 for all regions, i.e., the difference will increase as the average of CBFPET and CBFASL increases. There is also an apparent underestimation of CBFASL in subcortical regions, which may be caused by shorter than assumed T1 relaxation and earlier arrival of labeled blood compared to cortical regions [9,22]. Previous studies comparing ASL and 15 O-water PET have reported correlation coefficients in GM ranging from 0.26 to 0.81 [4,10,11,[26][27][28][29][30][31][32][33] (see Figure 4). In our study, we found correlations between CBFASL and CBFPET varying between 0.42 in pallidum and 0.83 in frontal cortex. Two of the previous studies performed arterial blood sampling and used an integrated PET/MR. Zhang et al. [11]. reported a correlation of 0.80 in GM and 0.61 to 0.87 in cortical-and subcortical regions comparing simultaneously acquired CBF measurements. Although the reported correlation coefficients agree with ours, Zhang et al. [11]. found generally higher values for CBFASL than for CBFPET, which is contrary to our results. However, Zhang et al. [11], used template-based attenuation correction, which might underestimate PET tracer uptake [14,34]. Puig et al. [4]. compared simultaneously acquired CBF measurements during rest in healthy subjects and found a correlation of 0.32. Correlations in cortical-and subcortical regions were reported using combined data from rest and altered perfusion states, so any comparisons to our study are hard to make. Both Zhang et al. [11] and Puig et al. [4] also performed a Bland-Altman analysis and reported a bias (LoAL and LoAU) of 15 (−5 and 25, width 30) and 0 (−15 and 15, width 30) mL/100 g/min in GM [4,11], respectively, compared to −15 (−45 and 16, width 61) mL/100 g/min in the present study. No regression analysis was performed in any of the two studies. Furthermore, Bland-Altman analysis of cortical and subcortical regions was omitted or included several altered perfusion states in the two above mentioned studies, so no further comparison to our results is possible except for whole GM. When using RLoAs, the width between the upper and lower LoAs is drastically narrowed, indicating that the negative relationship between the difference and the average of CBFPET and CBFASL has a high impact on data interpretation. In GM, the width of the LoAs went from 61 mL/100 g/min to 35 mL/100 g/min, LoAU and LoAL were 16 and −45 mL/100 g/min, respectively. RLoAs are uniform around the regression line; therefore the corresponding upper and lower limits were 18 and −18 mL/100 g/min, which is more comparable to the LoAs reported by other investigations.
Still, given the width of the RLoAs compared to the normal-range CBF values, the agreement between CBFASL and CBFPET is not sufficient to be used interchangeably for When using RLoAs, the width between the upper and lower LoAs is drastically narrowed, indicating that the negative relationship between the difference and the average of CBF PET and CBF ASL has a high impact on data interpretation. In GM, the width of the LoAs went from 61 mL/100 g/min to 35 mL/100 g/min, LoA U and LoA L were 16 and −45 mL/100 g/min, respectively. RLoAs are uniform around the regression line; therefore the corresponding upper and lower limits were 18 and −18 mL/100 g/min, which is more comparable to the LoAs reported by other investigations.
Still, given the width of the RLoAs compared to the normal-range CBF values, the agreement between CBF ASL and CBF PET is not sufficient to be used interchangeably for measuring absolute and comparable CBF. However, re-scaling of ASL values using the relation between CBF PET and CBF ASL found in the present work could be considered. As different re-scaling would be required in cortical and subcortical regions, this may not be a feasible way to proceed.
We report an average CBF PET in GM at 75 mL/100 g/min. In contrast, previous studies using 15 O-water PET with arterial sampling have reported CBF values in GM ranging from 37 to 67 mL/100 g/min in healthy subjects [11,28,[35][36][37][38][39][40][41]. Thus, current published normal-range CBF measured by 15 O-water PET with arterial blood sampling shows large variations [36,42]. A generally accepted and often-cited average normal whole-brain CBF value in younger adults is 50 mL/100 g/min [43]. Moreover, average whole-brain, GM and WM CBF values of 50, 80, and 20 mL/100 g/min in neurologically normal subjects were early established using the Kety-Schmidt method with intra-arterial injection [36,44,45], which is in line with the CBF PET values found in the current work. We acknowledge that our reported average CBF PET in GM are high compared to other investigations, however, during quality control of our data we found no technical explanation. Of note, in subjects/patients with high average CBF PET , we also found high CBF ASL . Thus, a physiological explanation cannot be ruled out, but appears to be unlikely.
CBF derived from ASL is inherently dependent on the sequence implantation, vendor, and quantification method used. Therefore, caution is advised when generalizing the results and conclusions found here. Other investigations have stressed the importance of the PLD for the quantification of CBF with ASL. In studies where ASL and 15 O-water PET were compared in patients with cerebrovascular diseases affecting the blood transit time, PLD appeared to be a critial parameter that can affect the results. However, all subjects and patients in this study are regarded to have a normal blood transit time. Hence we have used a PLD of 2000 ms as described by Alsop et al. [8].
We evaluated agreement in the normal CBF range. In addition to ten healthy volunteers, we included eight patients with epilepsy. However, interictal focal hypoperfusion is expected to have a negligible impact on our results since we used mostly large VOIs. Moreover, an unpaired t-test between patients with epilepsy and healthy subjects did not reveal any significant differences for any region (results not shown). A thorough investigation of potential differences was outside the scope of the current study and therefore not reported in detail.

Conclusions
Although a high correlation between CBF ASL and CBF PET was found, the agreement in absolute and comparable CBF values was not sufficient for ASL to be used interchangeably with 15 O-water PET.
Supplementary Materials: The following are available online at https://www.mdpi.com/article/10 .3390/diagnostics11050821/s1, Table S1: Acquisition parameters of MRI scans, Figure S1: Representative example of VOI definition for a healthy subject, Document S1: Calculation of regression-based limits of agreement (RLoA)-example grey matter.  Institutional Review Board Statement: The study was conducted according to the guidelines of the Declaration of Helsinki, and approved by the Regional Board of Medical Ethics in Uppsala (DNR 2015/187).