Efficient Visualization of Vascular Territories in theHuman Brain by Cycled Arterial Spin Labeling MRI

Matthias Gunther1,2*

Intracranial vascular territories are usually visualized with theuse of angiographic techniques. However, because it is difficultto visualize the distal vascular bed with any angiographic tech-nique, it is also difficult to infer the parenchymal borders ofvascular territories. Arterial spin labeling (ASL) MRI providesinformation on cerebral perfusion, and a regional ASL (regASL)approach offers the potential to visualize perfusion of the an-terior and posterior circulation separately. Current techniquesperform the labeling in each feeding artery as a separate ex-periment, which is very time-consuming. In this work a verytime-efficient regASL technique is presented that acquires allvascular territories within the same experiment. Images withdifferent combinations of labeled arteries are combined to de-lineate the vascular territory of interest. Five subjects wereexamined with a clinical 1.5T MR scanner. A sharp delineationof the middle cerebral artery (MCA) and posterior cerebral ar-tery (PCA) territories with whole-brain coverage was achievedin all subjects. The resulting signal-to-noise ratios (SNRs) forconventional regASL and the proposed cycled regASL were9.3 � 2.4 and 15.3 � 5.1, respectively. An optimized setup wasachieved by combining the regional labeling scheme with anefficient readout technique, which yielded a total measurementtime of 2 min for three vascular territories. Magn Reson Med56:671–675, 2006. © 2006 Wiley-Liss, Inc.

Key words: perfusion; vascular territory; arterial spin labeling;Hadamard; high efficiency

The intracranial vascular territories are usually assessed invivo with the use of angiographic techniques. However, itis difficult to infer the parenchymal borders of vascularterritories. Arterial spin labeling (ASL) is a well knowntechnique for noninvasive measurement of brain perfusion(1,2), and some ASL techniques allow an arbitrary choiceof the region where the water spins are tagged. Recently,several methods based on ASL were described (3–6) thatseparate the perfusion signal into specific vascular territo-ries. Vascular territory perfusion measurement (regionalASL (regASL)) is accomplished by creating prepared spinmagnetization (e.g., inverted or saturated) in a specificlarge arterial vessel, rather than all feeding blood vessels.As in conventional ASL, two data sets (one with (labelimage) and one without (control image) preparation) aretypically acquired in a downstream position.

However, common to all published methods is that theblood of each feeding artery is labeled within a separate

experiment. This leads to a rather long total measurementtime if multiple vascular territories, such as the left andright internal carotid arteries (ICAs) and basilar artery,have to be acquired (9–15 min in Refs. 3 and 4). A verysimple method to speed up the acquisition without sacri-ficing the signal-to-noise ratio (SNR) of the resulting per-fusion images is to use the same control image (i.e., ac-quired without labeling of blood in a vessel) for all vascu-lar territory measurements. If three different vascularterritories are targeted, only four (rather than six) data setshave to be acquired. This reduces the measurement timeby one-third. However, only two of the four data sets areused to calculate the perfusion signal of one territory. Amore efficient method in terms of SNR is to encode vascu-lar territories such that all four data sets will be used forextraction of the perfusion signal of a single territory.

This work presents a method that acquires all threeterritories in one experiment and yields the same SNR andmeasurement time as the previous methods for one terri-tory. To achieve this, a labeling scheme is used to tag theblood of two vessels simultaneously. The combination oflabeled vessels is varied throughout the experiment,which allows separation of each single territory. By usingan efficient readout technique (e.g., single-shot 3D-GRAdi-ent and Spin Echo imaging (GRASE) (7)), one could furtherreduce the total acquisition time to 2 min, which wouldallow routine use of the method in clinical protocols.

THEORY

An ASL experiment can be viewed as a kind of gradientcycle experiment in which the sign of the longitudinalspin magnetization of the inflowing blood changes be-tween �1 (control image) and –1 (label image) if relaxationis neglected. In conventional gradient cycling experimentsthe amplitude and/or direction of one or more gradientpulses is changed (“cycled”) over several repetitions. InASL experiments the extent and position of the inversion(or saturation) slab change between the control and labelimages. In this regard, ASL experiments are similar togradient cycle experiments with two phases. In this studythe ASL experiment consisted of four different phases, andthe respective inversion slabs covered different feedingvessels in each phase.

The position and orientation of the inversion slab foreach of these four phases are shown in Fig. 1. With thiscycled labeling scheme the vascular territory of three ves-sels (left and right ICAs and posterior circulation suppliedby the basilar artery) can be differentiated. The table in Fig.1 denotes the vessels in which the blood spins are labeled(“label”) or remain unchanged (“control”) in each phase.In phases 1–3 blood spins within two of the three vesselsare labeled, while they remain unchanged in the other one.

1Department of Neurology, Universitatsklinikum Mannheim, University of Hei-delberg, Heidelberg, Germany.2Advanced MRI Technologies, Sebastopol, California, USA.*Correspondence to: Dr. Matthias Gunther, Department of Neurology, Uni-versitatsklinikum Mannheim, University of Heidelberg, Theodor-Kutzer-Ufer1–3, 68167 Mannheim, Germany. E-mail: matthias.guenther@mediri.comReceived 4 January 2006; revised 2 May 2006; accepted 27 May 2006.DOI 10.1002/mrm.20998Published online 10 August 2006 in Wiley InterScience (www.interscience.wiley.com).

Magnetic Resonance in Medicine 56:671–675 (2006)

© 2006 Wiley-Liss, Inc. 671

The combination of labeled vessels changes with thephases. In phase 4 no labeling of blood spins is performedin all feeding vessels. Thus, this data set is used as acontrol set for all vascular territories.

By subtracting the data sets of any two different labelingphases, exactly two vascular territories can be distin-guished if there is no overlap in vascular territories other-

wise the signal would (partially) cancel out (this mightoccur if only two data sets are combined, by using all fourdata sets overlapping vascular territories can be distin-guished). This can best be appreciated by subtracting thedata set of phase 4 from the data acquired in phases 1–3(second row of Fig. 2). Each subtraction data set shows twodifferent combinations of exactly two vascular territories.

FIG. 1. Demonstration of the positioning and ori-entation of labeling slabs on a high-resolution 3Dtime-of-flight MRA data set. Both ICAs, the verte-bral arteries, the basilar artery, and the circle ofWillis are shown. Due to the proximity of the ves-sels of interest, careful slab positioning to avoidlabeling bleed-over is necessary. A four-pulse cy-cle is shown for separation of the left and rightICAs and basilar territory. [Color figure can beviewed in the online issue, which is available atwww.interscience.wiley.com.]

FIG. 2. Postprocessing step for reconstruction ofseparated vascular territories. [Color figure can beviewed in the online issue, which is available atwww.interscience.wiley.com.]

672 Gunther

One must use an even number of images to reconstructthe vascular territories, since otherwise the backgroundsignal of stationary tissue will not vanish. Therefore, it isadvantageous to acquire an additional set of data in case anodd number of regions have to be separated. The data set ofphase 4 is not essential for successful separation of thevascular territories, but it serves as a “stop-gap” that in-creases efficiency by allowing all four data sets to be usedfor reconstruction of each regional data set. The properlinear combinations of the labeling phases are shown inthe rows of the table in Fig. 1 (label � –1 and control ��1), i.e., to separate the left ICA region the sum of phases2 and 3 must be subtracted from the sum of phases 1 and4 (see Fig. 2).

MATERIALS AND METHODS

Five healthy human subjects were examined (mean age �38 � 13 years) on a clinical 1.5 T MR scanner (MagnetomSonata; Siemens, Erlangen, Germany) with maximum gra-dients of 40 mT/m and a minimum rise time to full gradi-ent strength of 200 �s. The procedure was performed ac-cording to the rules of the local ethics committee.

The goal of the study was to separate three vascularterritories (the left and right ICAs and basilar artery). Theinversion slabs were positioned on coronal and transverseprojections of a reconstructed 3D MR angiography (MRA)data set, as shown in Fig. 1. Precise localization is impor-tant to avoid bleed-over effects.

The MR sequence was implemented as shown in thediagram (Fig. 3). Only the gradient pulses that define theorientation of the slice-selective inversion RF pulse (label-ing pulse) are unique for each cycle phase, as defined inFig. 1. Adiabatic RF pulses (hyperbolic secant) were em-ployed for inversion. Efficient postlabeling saturation ofthe imaging slab was realized by a four-RF-pulse train asdescribed in Ref. 4. This is necessary to ensure properspoiling of the effect of any labeling pulse within theimaging slab. These effects include direct labeling of statictissue, as well as differences caused by magnetizationtransfer. In addition, two nonselective adiabatic hyper-bolic secant pulses were used for background suppressionas described in Ref. 7.

A single-shot 3D-GRASE readout technique (7) was usedwith the following parameters: inflow time (TI) � 1500 ms,echo time (TE) � 36 ms, repetition time (TR) � 2500 ms,off-resonance fat saturation pulse, 26 interpolated parti-tions (16 acquired, 5/8 Fourier), FOV � 300 mm �150 mm � 117 mm, and matrix size � 64 � 33 (interpo-lated to 128 � 64). A near-isotropic resolution of 4.7 mm �4.7 mm � 4.5 mm was achieved. Because of the relatively

high SNR of the single-shot 3D acquisition, five repetitionswere sufficient for a total acquisition time of only 2 min.

The postprocessing procedure is illustrated in Fig. 2.After data acquisition the data sets were combined asshown in Fig. 1 to yield the separated vascular territory ofeach particular feeding vessel. Data sets in which bloodspins of a particular feeding vessel (left/right ICA or basilarartery) were labeled were added, while the control phaseswere subtracted. Intensity images of all three vascularregions were combined into a single image by using threeindependent color channels (red, green, and blue).

An independent component analysis using the SHIftedBlocks for Blind Separation (SHIBBS) algorithm (8) wasperformed using the four acquired data sets. This algo-rithm optimizes a fourth-order measure of independencealgebraically under the whiteness constraint. This ap-proach allows for parameter-free estimation of indepen-dent components.

For comparison, conventional regASL measurementswere performed for each of the three feeding arteries. Thelabeling slabs were positioned as described in Refs. 3 and4. The same control data set was used for all three vascularregions. The same acquisition parameters as describedabove were applied, for a total measurement time of 2 minfor all three vascular territories.

SNR measurements were obtained by region-of-interest(ROI) analysis (i.e., the mean intensity of a tissue ROI ineach vascular region was divided by an ROI in air). ROIswere preferably selected in the four central slices of theseparated vascular territory perfusion data set (bottom rowin Fig. 2).

RESULTS

Colored rCBF maps demonstrated the cortical vascularperfusion pattern and provided a sharp delineation ofanterior and posterior vascular territories in all five sub-jects. In two of the five subjects only partial separation ofvascular territories could be achieved, presumably due tomotion between the localization scan and the regASL mea-surement. The SNRs (� standard deviation (SD)) of allsubjects for conventional and cycled regASL were 9.3 �2.4 and 15.3 � 5.1, respectively. Figure 4 shows the vas-cular territory mapping for one subject with asymmetricICA territories using conventional regASL (Fig. 4a) and theproposed method of cycled regASL before (Fig. 4b) andafter (Fig. 4c) independent component analysis. The leftICA territory is shown in green, the right ICA territory isred, and the basilar artery territory is blue. Figure 5 showsmultiplanar reconstructions of the cycled regASL data setof another subject with symmetric ICA territories without

FIG. 3. Sequence diagram for effi-cient regASL for a full four-phasecycle. Please note that parts of thesequence (Water Suppression En-hanced through T1 effects (WET)pulse and background suppressionpulses) were omitted in phases 2–4for better visualization.

Vascular Territories by Cycled ASL 673

independent component analysis. The resulting data set wasmedian filtered (isotropic 2 pixel) to improve display.

The resulting vascular distinctions in all subjects ex-actly matched the known anatomical vascular supply bor-ders (e.g., along the thalamus) (9).

The independent component analysis did not changethe result significantly in three subjects. In two subjectsthe insufficient separation of vascular territories was im-proved (see Fig. 4b and c).

DISCUSSION

Conventional ASL (including prior vascular territorywork) can be viewed as a gradient cycle experiment with

two different phases (control and label images). As a moreefficient way to acquire more than one vascular territory,we developed a four-phase cycle experiment. This methodincurs no time penalty, since several repetitions have to beperformed in standard ASL measurements to compensatefor poor SNR.

By selectively labeling blood spins in certain vessels, theparenchymal vascular territory of the vessels can be visu-alized. Several different approaches can be used, but theyall have in common that each vascular territory is handledseparately. This leads to very time-consuming experi-ments that are not easy to integrate into existing clinicalprotocols. The goal of our work was to develop a moreefficient technique that would allow fast acquisition of atleast three vascular territories (left/right ICA and basilarartery). This was achieved by combining the measurementof all vascular regions in one experiment in an SNR-effi-cient way as described above. An improved single-shot 3Dreadout technique was used, which increased the resultingSNR and allowed almost whole-brain coverage. Linearcombinations of the data sets of the three vascular regions(left/right ICA and basilar artery) resulted in a data set withthe same quality and SNR as acquired with conventionalASL techniques (data not shown). Therefore, most perfu-sion measurements performed with conventional ASLcould be performed with regASL in the same measurementtime without sacrificing quality. This may also hold truefor more complex ASL experiments, such as time-seriesacquisitions, if at least one repetition per time step isperformed. Furthermore, cycled regASL provides addi-

FIG. 4. Comparison of (a) conventional regASL, (b) cycled ASL, and(c) cycled ASL with independent component analysis on a 36-year-old subject with asymmetric ICA territories. Images a and c showcomparable visualization of vascular territories, including the clearasymmetry in the left and right ICA territories. However, a lower levelof noise can be seen in images b and c due to the higher acquisitionefficiency of cycled ASL. Incomplete separation (less brilliant colorsaturation in b) due to motion between acquisitions of the localizerscan and the regASL data can be improved by applying an inde-pendent component analysis to the cycled ASL data set.

FIG. 5. Transverse, coronal, and sagittal views of a median filtered 3D dataset from a 38-year-old subject with symmetric ICA territories. Color codingdelineates the tissues of the different vascular territories (green � left ICA,red � right ICA, blue � basilar artery). Note the close proximity of blue andgreen on the lateral border of left thalamus, which exactly matches theanatomical vascular supply territories in this border zone region. Acquisi-tion time for the whole data set was 2 min.

674 Gunther

tional information on the vascular territories of at leastthree major vessel subtrees. However, no correspondingexperiments have been performed yet.

In general, N regions can be distinguished by N � 1 datasets with the proposed technique, while there is need for atleast 2 � N data sets in the conventional technique, or only N� 1 if the same control data set is used for all regions (opti-mized conventional regASL). The main improvement of theproposed technique (cycled regASL) compared to conven-tional regASL is the increased SNR. By using proper linearcombinations of labeled and unlabeled regions (Hadamardencoding), an optimized SNR in cycled regASL can be ob-tained, since all (N � 1) data sets (instead of just two datasets, as in conventional regASL) are used to calculate eachvascular region. Thus, SNR will be higher by a factor of�(2 � N)/�(2) � �(N) compared to standard techniques(3,4), and �(N � 1)/�(2) for optimized conventional regASLwith only one control data set as used in the describedexperiment. Strictly speaking, the SNR calculation is onlyvalid for non-overlapping vascular regions, since otherwisethe noise might be correlated. The measured ratio of SNR incycled regASL to conventional regASL was 1.65 � 0.47,which is close to the expected value of �(2).

Further improvement by using a higher-resolution label-ing scheme utilizing a grid of labeled areas may provide anopportunity to separate vascular territories in a postpro-cessing analysis. Such approaches have been used forhigh-resolution imaging (10), but have not yet been ap-plied to labeling pulses and ASL. The spatial definition ofthe labeling slab to create a grid can be determined byusing Walsh functions. Walsh functions are definingeigenfunctions with function values of �1 and –1 only.Here, �1 represents no labeling, while –1 represents label-ing of the corresponding region. The separately acquireddata sets can then be combined to yield the regional infor-mation using corresponding linear combinations as pro-posed in this work. This type of Hadamard encoding re-quires the total number of measurements to be a power of2. If N � 1 is not 2, 4, 8, . . ., no loss of SNR is expected aslong as the number of acquisitions is a multiple of the nextpower of 2 (smallest k that fulfills N � 1 � 2k).

Careful positioning of the labeling slabs on a high-resolu-tion 3D MRA data set is an important prerequisite, and re-quires expert anatomical knowledge (otherwise, nonoptimalseparation of the vascular territories may result). Indepen-dent component analysis improved separation of vascularterritories in the case of motion between acquisition of thelocalizer images and measurement of the cycled regASL data.However, it is not clear how this algorithm performs foroverlapping vascular territories. Further research is required

to assess the performance of independent component analy-sis in this application in more detail.

The orientation of the labeling slabs was similar to thosedescribed in the literature (3,4). However, the position issomewhat complementary since each labeling slab in theproposed technique covers two vessels, rather than justone vessel. For example, one labeling slab in conventionalregASL is supposed to label the right ICA only (and notlabel anything else), whereas the labeling slab in ourmethod is supposed to label vessels other than the rightICA (i.e., the left ICA and vertebra/basilar artery). In thissense the positioning of the labeling pulses in conven-tional regASL and our method is comparable.

CONCLUSIONS

The current results show that improved visualization of theparenchymal borders of vascular territories is feasible withcycled regASL. Information on the vascular origin of localperfusion can be achieved. The proposed technique providessuch information in a more efficient way than conventionalregASL techniques. In combination with an SNR-efficientsingle-shot 3D readout strategy, a clinically useful data setcan be acquired in a few minutes. This is a first step towardobtaining more detailed information on regional brain perfu-sion.

REFERENCES1. Wong EC, Buxton RB, Frank LR. Quantitative imaging of perfusion

using a single subtraction (QUIPSS and QUIPSS II). Magn Reson Med1998;39:702–708.

2. Kim SG. Quantification of relative cerebral blood flow change by flow-sensitive alternating inversion recovery (FAIR) technique: applicationto functional mapping. Magn Reson Med 1995;34:293–301.

3. Hendrikse J, van der Grond J, Lu H, van Zijl PC, Golay X. Flow territorymapping of the cerebral arteries with regional perfusion MRI. Stroke2004;35:882–887.

4. Golay X, Petersen ET, Hui F. Pulsed star labeling of arterial regions(PULSAR): a robust regional perfusion technique for high field imaging.Magn Reson Med 2004;53:15–21.

5. Eastwood JD, Holder CA, Hudgins PA, Song AW. Magnetic resonanceimaging with lateralized arterial spin labeling. Magn Reson Imaging2002;20:583–586.

6. Davies NP, Jezzard P. Selective arterial spin labeling (SASL): perfusionterritory mapping of selected feeding arteries tagged using two-dimen-sional radiofrequency pulses. Magn Reson Med 2003;49:1133–1142.

7. Gunther M, Oshio K, Feinberg DA. Efficient 3D perfusion measurementusing single-shot 3D-GRASE. Magn Reson Med 2005;54:491–498.

8. Cardoso JF. High-order contrasts for independent component analysis.Neural Comput 1999;11:157–192.

9. Hanaway J, Woolsey T, Gado M, Roberts M. The brain atlas: a visualguide to the human central nervous system. Bethesda, MD: FitzgeraldScience Press; 1998.

10. Fletcher DW, Haselgrove JC, Bolinger L. High-resolution imaging usingHadamard encoding. Magn Reson Imaging 1999;17:1457–1468.

Vascular Territories by Cycled ASL 675