the effect of varying user-selected input parameters on quantitative values in ct perfusion maps1

Original Investigations

The Effect of Varying User-Selected InputParameters on Quantitative Values in CT

Perfusion Maps1

Pina C. Sanelli, MD, Michael H. Lev, MD, James D. Eastwood, MD, R. Gilberto Gonzalez, MD, PhD, Ting Y. Lee, PhD

Rationale and Objectives. Deconvolution-based software can be used to calculate quantitative maps of cerebral bloodflow (CBF), cerebral blood volume (CBV), and mean transit time (MTT) from first-pass computed tomography perfusion(CTP) datasets. The application of this software requires the user to select multiple input variables. The purpose of thisstudy was to investigate the degree to which both major and minor variations of these user-defined inputs would affect thefinal quantitative values of CBF, CBV, and MTT.

Materials and Methods. A neuroradiologist constructed CBF, CBV, and MTT maps using standard methodology withcommercially available software (GE Functool Version 1.9s) from CTP datasets of three acute stroke patients. Each mapwas reconstructed multiple times by systematically and independently varying the following parameters: postenhancementand preenhancement cutoff values, arterial and venous region-of-interest (ROI) placement, and arterial and venous ROIsize. The resulting quantitative CTP values were compared using identical ROIs placed at the infarct core.

Results. Major variations of either arterial ROI placement or arterial and venous ROI size had no significant effect on themean CBF, CBV, and MTT values at the infarct core (p � .05). Even minor variations, however, in the choice of venousROI placement or in pre- and postenhancement cutoff values significantly altered the quantitative values for each of theCTP maps, by as much as threefold.

Conclusion. Even minor variations of user-defined inputs can significantly influence the quantitative, deconvolution-based CTPmap values of acute stroke patients. Although quantitation was robust to the choice of arterial ROI placement and arterial orvenous ROI size, it was strongly dependent on the choice of venous ROI location and pre- and postenhancement cut-off values.Awareness of these results by clinicians may be important in the creation of quantitatively accurate CTP maps.

Key Words. CT perfusion; quantitative analysis; acute stroke.© AUR, 2004

An important goal of stroke imaging—which can potentiallybe accomplished using perfusion scanning—is to accurately

Acad Radiol 2004; 11:1085–1092

1 From the Department of Radiology, Weill Medical College of Cornell Univer-sity, New York Presbyterian Hospital, 520 East 68th Street, Starr Pavilion,Starr-630, New York, NY 10021 (P.C.S.); Department of Radiology, HarvardMedical School, Massachusetts General Hospital, Boston, MA (M.H.L.,R.G.G.); Department of Radiology, Duke University Medical Center,Durham, NC (J.D.E.); and John P. Robarts Research Institution, London,ON, Canada (T.Y.L.). Received February 2, 2004; revision received and ac-cepted June 6, 2004. Address correspondence to P.C.S. e-mail:[email protected]

©
AUR, 2004doi:10.1016/j.acra.2004.07.002
distinguish infarct “core” (nonviable tissue) from “ischemicpenumbra” (surrounding tissue that is still viable, but at riskfor infarction). In the absence of early recanalization, a pa-tient with a core-penumbra mismatch will typically exhibitgrowth of the infarct core into the region of the ischemicpenumbra. Hence the role of thrombolytic therapy is to pre-vent extension of infarction by early restoration of bloodflow to the penumbra. Given the potential for hemorrhageand the subsequent narrow time window for treatment, care-ful patient selection is critical.

Evaluation of cerebral perfusion has become an impor-tant step in determining the presence and extent of the
infarct core and ischemic penumbra. This information can
1085

SANELLI ET AL Academic Radiology, Vol 11, No 10, October 2004

be obtained from magnetic resonance imaging (MRI) orcomputed tomography (CT) imaging (1–3). Given thelimitations of MRI, including its restricted availability,lengthy exam duration, and patient compatibility issues,CT perfusion (CTP) has emerged as a promising modal-ity. Because unenhanced CT scan of the head is routinelyperformed before thrombolysis to exclude intraparenchy-mal hemorrhage, CT angiography with quantitative CTPcan be readily added to the diagnostic armamentariumimmediately following this, without significant delays indefinitive treatment (4–7).

CT perfusion data are obtained by a cine acquisition dur-ing the dynamic administration of a 40–50 mL contrast bo-lus. Of note, this limits each CTP acquisition to a 2-cm slabof coverage (the detector length of most CT scanners).When a numerical deconvolution model is used for analysis,truly quantitative maps of cerebral blood flow (CBF), cere-bral blood volume (CBV), and mean transit time (MTT) canbe constructed from this CTP dataset (7). Such a model re-quires the selection of a region of interest (ROI) from anartery and vein in the field-of-view of the image to providerepresentative arterial and venous time-attenuation curves(TAC) reflecting the “arterial input” and “venous outflow”from the parenchymal capillary network. This arterial inputfunction is required mathematically to perform the deconvo-lution. The venous curve is required for quantitation to cor-rect the arterial input for volume averaging effects. In addi-tion to selecting arterial and venous ROI size and location,the observer also needs to define where along the time-axisof the TAC to start and end the deconvolution calculation—the so called “pre-” and “post-” enhancement cutoff values.

Theoretical guidelines already exist as to the optimal se-lection of these user-defined input variables for the creationof accurate, qualitative CBV, CBF, and MTT maps (7). Thepurpose of this article is to confirm and underscore the de-gree to which both major and minor variations of these user-defined inputs affect the final, quantitative values of CBF,CBV, and MTT, and hence to provide practical pointers forthe construction of quantitative CTP maps using the decon-volution method.

MATERIALS AND METHODS

CT perfusion datasets from three patients were ana-lyzed in this study. All patients presented with acute mid-dle cerebral artery infarction; two on the right side andone on the left side. Two male and one female patients
were included in the study. The age range was 30–63
1086

years of age. These three datasets were acquired at twodifferent institutions on GE HiSpeed and LightSpeed CTscanners (General Electric Medical Systems, Milwaukee,WI). The protocol for performing CT perfusion used cinemode (acquisition of one image per second) with kVp/mAset at 80/190; power injection of 45 cc of nonionic con-trast at 4.0–5.0 cc/sec with a 5-second delay. The gantryangle was selected to avoid radiation exposure to the or-bits. The total scan time was 45 seconds. The slice selec-tion was chosen at the level of the basal ganglia to in-clude vascular territory from the anterior, middle, andposterior cerebral arteries. The acquired CT perfusionimages were transferred to a GE Advantage Windowsworkstation (General Electric Medical Systems, Milwau-kee, WI) for postprocessing of CBF, CBV, and MTTmaps.

Using the same software program (GE Functool ver-sion 1.9s), a single neuroradiologist constructed multiplequantitative CBF, CBV, and MTT maps for all three pa-tients. The GE Advantage Windows workstation resourcefile parameters were set at standard values, with “minT-ransitTime” � 1.0, “maxBloodFlow” � 400, and “max-TimeResolution” � 1.0. Sets of CBF, CBV, and MTTmaps were repeatedly constructed for each patient byvarying a single user-defined parameter, with the remain-ing parameters held constant, as described in the follow-ing section.

Six user-defined parameters were chosen for study(only four for the MTT maps), and assigned four suc-cessive values each, for a total of 24 CBF, 24 CBV,and 16 MTT maps, or 64 maps per patient. Each pa-rameter was varied in turn, with the default values forthe parameters held constant being those that were rec-ommended by the user manual for the GE Functoolversion 1.9s.

The six user-defined parameters that were studiedwere: (1, 2) the location of the ROI selected for thearterial and venous TACs; (3, 4) the size of the ROIselected for the arterial and venous TACs; and (5, 6)the pre- and post enhancement “cut-off time points”selected for the arterial and venous TACs (ie, where tostart and stop the deconvolution calculations). The fourselections that were made per parameter are delineatedin the captions of Figs. 1–9, as follows: (1) the posten-hancement cut-off time points were based on the imagenumber following the initial down-slope of the arterial,venous and tissue TACs (Fig. 1); (2) the preenhance-ment cut-off time-points were based on the image num-
ber preceding the initial up-slope of the arterial and

are shown.

tion choices (arrows 1–4) are shown.

Academic Radiology, Vol 11, No 10, October 2004 INPUT PARAMETERS ON QUANTITATIVE VALUES

venous TACs (Fig. 2); (3) the ROI locations for thearterial and venous TACs (Fig. 3–5) were selected, se-quentially, from the largest to the smallest feeding ves-sels ipsilateral and contralateral to the infarction; and(4) the ROI sizes for the arterial and venous TACswere selected, sequentially, from smallest to largest,ranging in area from 1 to 16 square pixels (Fig. 6, 7).

The multiple CBV, CBF, and MTT maps that werecreated for each patient were analyzed by sampling andcomparing the quantitative values of each parameter,on each map, at the infarct “core” and unaffected grayand white matter. The unaffected gray and white matterwas sampled from the contralateral normal hemisphere.“Core” was defined by the location of the maximal le-sion abnormality on the CT, CTP, and diffusion-weighted MRIs. The ROIs sampled from the core, andunaffected gray and white matter for each of the mapswere identical in both size and placement, so as toeliminate a potential source of error in the comparisonbetween the different maps. The quantitative mean ROI

Figure 3. Axial CT source image from computed tomographyperfusion dataset. Marker 1 shows selection of a representative,sample “arterial input” region of interest (ROI), located at the dis-tal internal carotid artery (ICA). Marker 2 shows selection of a rep-resentative “venous outflow” ROI, located at the torcula.

Figure 1. Arterial (small #1), venous (small #2), and tissue(small #3) time-attenuation curves generated for computationof perfusion maps from the sites in Fig. 3. The selectedpostenhancement variable cutoff location choices (arrows 1– 4)

Figure 2. Arterial (small #1) and venous (small #2) time-attenua-tion curves generated for computation of perfusion maps from thesites in Fig. 3. The selected preenhancement variable cutoff loca-

values for the CBF, CBV, and MTT maps from each of

1087


the three patients, for each of the experimental vari-ables, were graphed and analyzed visually. The x-axisof each graph in the results section (as defined in Fig.4 –9) reflects the variable, sequential values selected forthe parameter under study. Mean values were compared

Figure 4. Summary graphs for all patients varying the arterialregion of interest (ROI) location used to select the arterial time-attenuation curve for computation of the computed tomographyperfusion maps. There is no significant difference in the quantita-tive values of cerebral blood flow, cerebral blood volume, andmean transit time for the sampled infarct core. Therefore, choos-ing an intracranial arterial vessel has no restrictions, whether it isa large supplying vessel ipsilateral to the infarction or a smallcontralateral vessel. ROI locations on the x-axis correspond todistal internal carotid artery (ICA), ipsilateral middle cerebral artery(MCA) to the infarction, ipsilateral anterior cerebral artery (ACA),and contralateral ACA vessels, respectively.

using the Student’s t-test.

1088

RESULTS

All three patients had acute middle cerebral artery ter-ritory infarctions identified on MR diffusion-weightedimaging. A total of 192 CTP maps were constructed (64maps per patient).

Overall, the changes in the CTP quantitative valueswere most significant in the infarct core compared withthe unaffected gray and white matter sampled from thecontralateral side. The unaffected gray and white matterdid not necessarily reflect these changes, likely because oftheir higher quantitative values. In this study, emphasishas been placed on analyzing the infarct core for changesin the quantitative values because it is the most well stud-

Figure 5. Summary graphs for all patients varying the venousregion of interest (ROI) location used to select the venous time-attenuation curve for computation of the computed tomographyperfusion maps. Increased cerebral blood flow and cerebral bloodvolume quantitative values occurs when using smaller deep ve-nous structures (locations 3 and 4) for the sampled infarct core,with as much as a threefold difference. These locations representdeep venous drainage structures, such as the vein of Galen andinternal cerebral veins. Deep venous structures only drain a lim-ited territory of brain and are not considered representative of theentire region studied. Computation of mean transit time mapsdoes not require a venous curve, and therefore were not calcu-lated. ROI locations on the x-axis correspond to the center of thetorcula, lateral torcula, vein of Galen, and internal cerebral veins,respectively.

ied region, with good evidence that visually apparent le-


sions reflect tissue likely to be irreversibly infarcted. Mostimportantly, it is discrepancies in the core values that aremost likely to affect clinical decision making.

As shown in Fig. 6 and 7, the mean values of CBF,CBV, and MTT obtained at the infarct core were inde-pendent of the size of the ROIs selected for the arterialand venous TACs (P � .05). Similarly, the mean CBF,

Figure 6. Summary graphs for all patients varying the arterialregion of interest (ROI) size used to select the arterial time-attenu-ation curves for computation of the computed tomography perfu-sion maps. There is no significant difference in the quantitativevalues of cerebral blood flow, cerebral blood volume, and meantransit time for the sampled infarct core. The arterial ROI size hasno effect as long as it is sampling within the lumen of the vessel.ROI sizes on the x-axis represent pixel areas of 1, 4, 9, and 16,respectively.

CBV, and MTT values at the infarct core were indepen-

dent of the location of the ROI selected for the arterialTAC (P � .05). In other words, choosing a large proxi-mal intracranial artery or a smaller distal artery ipsilateralor contralateral to the infarct did not affect the resultingquantitative perfusion values. Figure 4 summarizes theseresults.

For the location of the ROI selected for the venousTAC, however, there was statistically significant differ-ence between the mean quantitative CBF and CBV mea-surements obtained at the infarct core (the computation ofMTT maps does not require a venous input parameter, sothis was not examined). Sampling of the smaller, deepvenous structures resulted in increased CBF and CBVvalues for the infarct core; the CBF and CBV quantitativemeasurements were increased by a factor of three whenusing a smaller deep venous structure, such as the internalcerebral vein, compared with the superior sagittal sinus ortorcula. The percent difference in the mean quantitative

Figure 7. Summary graphs for all patients varying the venousregion of interest (ROI) size used to select the venous time-attenuation curve for computation of the computed tomogra-phy perfusion maps. There is no significant difference in thequantitative values of cerebral blood flow and cerebral bloodvolume for the sampled infarct core. The venous ROI size hasno effect as long as it is sampling within the lumen of the ves-sel. Computation of mean transit time maps does not require avenous input. ROI sizes on the x-axis represent pixel areas of1, 4, 9, and 16, respectively.

values for the infarct core was 199% for CBF and 258%

1089


for CBV. However, the overall mean CBF remained�10 mL/100 g/min, and CBV �3 mL/100 g, at the in-farct core, independent of the choice of venous ROI loca-tion. Figure 5 summarizes these results.

The time point chosen for the postenhancement cut-off

Figure 8. Summary graphs for all patients varying the posten-hancement parameter for computation of the computed tomogra-phy (CT) perfusion maps. Increased cerebral blood flow, de-creased cerebral blood volume, and shortened mean transit timevalues occur when using the postenhancement cut-off parameterimmediately after the initial downslope on the baseline of the arte-rial time-attenuation curve (location 4) for the sampled infarctcore. This location eliminates from the computation part of thevenous curve; which is used to compute a scaling factor to cor-rect for partial volume averaging in generation of the CT perfusionmaps. Postenhancement variables on the x-axis correspond toFig. 1.

parameter also affected the resulting mean quantitative

1090

CBV, CBF, and MTT values at the infarct core, as sum-marized in Fig. 8. When a cut-off at the TAC value im-mediately following the arterial down-slope was used forthe deconvolution calculations (which included the mid-portion of the venous up-slope curve), there was in-creased CBF, decreased CBV, and shortened MTT at the

Figure 9. Summary graphs for all patients varying the preen-hancement parameter for computation of the computed tomogra-phy (CT) perfusion maps. Increased cerebral blood flow and cere-bral blood volume with shortened mean transit time occur whenusing the first image variable (location 1) for the sampled infarctcore. This location includes all of the nonenhancing images be-fore contrast injection, which prolongs the time course of the dataset and introduces increased noise and error into the generatedCT perfusion maps. Preenhancement variables on the x-axis cor-respond to Fig. 2.

infarct core. The percent difference in the mean quantita-


tive values for the infarct core was 34% for CBF, 51%for CBV, and 60% for MTT. However, the overall meaninfarct core values remained �10 mL/100 g/min for CBF,�3 mL/100 g for CBV, and �6 sec for MTT, despite theprecise choice of postenhancement cut-off time point.

Similarly, the time point chosen for the preenhance-ment cut-off parameter also had a significant impact onthe quantitative CTP values at the infarct core (Fig. 9).The percent difference in the mean quantitative values forthe infarct core was 51% for CBF, 34% for CBV, and53% for MTT. However, at the infarct core, the meanCBF remained �10 mL/100 g/min, mean CBV �3 mL/100 g, and mean MTT �6 sec, despite choice of preen-hancement cut-off time point.

DISCUSSION

In this article, we have confirmed and underscored thetheoretical prediction that major variations of either arte-rial ROI placement or arterial and venous ROI size hadno significant effect on the mean CBF, CBV, and MTTvalues at the infarct core (P � .05). Even minor varia-tions, however, in the choice of venous ROI placement,or in pre- and postenhancement cut-off time point valuessignificantly altered the quantitative values for each of thedeconvolution based CTP maps by as much as threefold.

The construction of quantitative CTP perfusion mapsrequires a deconvolution-based CT tracer kinetic model.Unlike MRI-based perfusion maps, the relationship be-tween tissue attenuation, measured in Hounsfield units,and the concentration of iodinated contrast material in thebrain capillaries, is linear (6). The software program forcreating quantitative CTP maps is not fully automated,and requires an observer to input specific user-definedparameters, thereby introducing variability and potentialerror in the construction of quantitative CT perfusionmaps.

The initial step in constructing CTP maps requires se-lection of an arterial and venous TAC. In our study, wevaried the intracranial arterial TAC by sampling from thelargest vessel perpendicular to the imaging section (usu-ally the top of the internal carotid artery or proximal mid-dle cerebral artery segment), to the smallest distal vessels(M2 and A2 segments). In addition, we sampled both theartery supplying the infarcted territory (eg, ipsilateralmiddle cerebral artery branches) as well as nonsupplyingarteries (eg, contralateral middle cerebral artery, ipsilateral
anterior cerebral artery branches). The location of the ar-
terial input did not affect the quantitative measurements atthe infarct core for any of the CBF, CBV, and MTT maps(Fig. 4). Therefore, choosing an intracranial arterial vesselhas no restriction; however, it should be noted that we atall times avoided choosing an ROI size larger than thevessel lumen diameter so as to minimize partial volumeaveraging. As long as there is no volume averaging inselecting the arterial input function, the GE software pro-gram is able to correct for the variation in the arrival timeof the contrast among the different vessels used. In addi-tion, the venous curve is used to determine a scaling fac-tor for the arterial input to account for partial volume av-eraging if it exists. One of the fundamental assumptionsin the quantification of perfusion data is the absence ofdelay and dispersion of the bolus between the site wherethe arterial input function is measured and the most dis-tant tissue; that is, the model assumes that the arterialinput function reflects the exact input to the tissue (8). Ifthere is a delay in the arrival of the contrast agent to thetissue from the arterial sampling location, then the decon-volved tissue impulse residue curve is shifted in time (9)in the GE CT perfusion software. Because shifting thedeconvolved impulse residue curve does not change thearea (CBV) and height (CBF) of the impulse residuecurve, the parameter values of CBF, CBV, and MTT re-main unchanged. The dispersion effect is very minimal asis evidenced by our results.

Indeed, in the deconvolution software program for con-structing CTP maps, the venous TAC is used to computea scaling factor for correcting for any partial volume av-eraging introduced by the choice of arterial input functionROI. At the level of the torcula, the draining venous si-nuses are large and perpendicular to the imaging plane. Inour study, choice of the venous ROI location did signifi-cantly influence the quantitative values of CBF and CBVat the infarct core. Using smaller vessels, such as the in-ternal cerebral veins and straight sinus/vein of Galen, re-sulted in partial volume averaging and hence resulted inelevated CBF and CBV values (Fig. 5).

The pre- and postenhancement cut-off time points areused to determine the start and end points on the TAC forthe computation of CTP maps by the deconvolutionmethod, thereby influencing the time needed to calculateperfusion maps and more importantly, the associated ac-curacy. If recirculation of contrast occurs, the arterial andvenous TACs may reveal a smaller, second peak follow-ing the initial large peak. Therefore, the postenhancementcut-off value should be at a time point after the initial
arterial downslope, when the TAC has reached a plateau,
1091


but before the second smaller upslope reflecting recircula-tion. Choosing the postenhancement cut-off value in thismanner will ensure that all or at least a significant down-slope of the venous TAC is included in the computation.Not all of the cases in our study had a recirculation peak.The reason that variations in the postenhancement timepoint values may influence quantitative CTP values at theinfarct core again relates back to the scaling function ofthe venous TAC. Because the venous curve is used tonormalize the arterial input function, to control for partialvalue averaging effects, truncation of the venous curvewill cause error in quantitation. Inclusion of very earlytime points may result in inaccuracies because of in-creased noise of the venous curve through inclusion ofunnecessary time points.

CONCLUSION

To limit potential quantitative error in constructingdeconvolution-based CT perfusion maps, we recommendusing a standardized procedure for the selection of theuser-defined parameters that are required to calculate theCBV, CBF, and MTT maps. Two specific guidelines forconstructing CT perfusion maps include: (1) avoid partialvolume averaging by selecting the largest vessel perpen-dicular to the imaging plane for the arterial and venousROIs, choosing an ROI pixel size to fit the lumen of thearterial and venous structures, and avoiding the use ofsmall, deep veins for the venous ROI; and (2) limit trun-cation of the TACs by selecting the postenhancement cut-off at a time point after the venous TAC has completedits downslope and reached a plateau (but before recircula-tion occurs) and by selecting the preenhancement cut-offat a time point immediately preceding the upslope of thearterial TAC.

In summary, we have confirmed and underscored thateven minor variations of certain user-defined inputs cansignificantly influence the quantitative CTP map values of
acute stroke patients. Specifically, although quantitation
1092

was robust to the choice of arterial ROI placement andarterial or venous ROI size, it was strongly dependent onthe choice of venous ROI location, as well as pre- andpostenhancement cut-off time points. Overall mean perfu-sion values at the infarct core, however, remained �10mL/100 g/min for CBF, �3 mL/100 g for CBV, and �6sec for MTT, despite the choice of input parameters. Typ-ically, there is an abrupt transition from infarct core andpenumbra tissue in these patients. If the quantitative val-ues are close to the borderline zone, it may affect the cat-egorization of the infarcted tissue. However, grossly thecategory did not change with small variations in the quan-titative values of CBF, CBV, and MTT. These values allremained below the threshold for the infarcted core. Tis-sue was not broadly misclassified in this small samplesize. Identification of infarct core using quantitative CTPthresholds, therefore, is highly reproducible, despite varia-tions in user-defined input parameters.

REFERENCES

1. Schlaug G, Benfield A, Baird AE, et al. The ischemic penumbra opera-tionally defined by diffusion and perfusion MRI. Neurology 1999;23:1528–1537.

2. Muzielaar JP, Fatouros PP, Schroder ML. A new method for quantitativeregional cerebral blood volume measurements using computed tomog-raphy. Stroke 1997; 28:1998–2005.

3. Mayer TE, Hamann GF, Wiesmann M, Brueckmann H, Klotz E. DynamicCT perfusion imaging in hyperacute stroke. Stroke 1998; 29:310.

4. Lev MH, Segal AZ, Farkas J, et al. Utility of perfusion weighted CT imagingin acute MCA stroke treated with intra-arterial thrombolysis: prediction offinal infarct volume and clinical outcome. Stroke 2001: 32;2021–2028.

5. Meier P, Zieler KL. On the theory of the indicator-dilution method formeasurement of blood flow and volume. J Appl Physiol 1954; 6:731–744.

6. Koenig M, Klotz E, Luka B, et al. Perfusion CT of the brain: diagnostic ap-proach for early detection of ischemic stroke. Radiology 1998; 209:85–93.

7. Nabavi DG, Cenic A, Craen RA, et al. CT assessment of cerebralperfusion: experimental validation and initial clinical experience.Radiology 1999; 213:141–149.

8. Calamante F, Gadian DG, Connelly A. Delay and dispersion effects indynamic susceptibility contrast MRI: simulations using singular valuedecomposition. Magnet Reson Medi 2000; 44:466–473.

9. Alsop DC. Perfusion MR imaging. In: Atlas SW, ed. Magnetic resonanceimaging of the brain and spine, 3rd ed. Philadelphia, PA: Lippincott
Williams & Wilkins, 2002; 215–238.

the effect of varying user-selected input parameters on quantitative values in ct perfusion maps1

Documents