Comparison of Errors Associated With Single- and Multi-Bolus Injection Protocols in Low-Temporal-ResolutionDynamic Contrast-Enhanced Tracer Kinetic Analysis

Caleb Roberts,* David L. Buckley, and Geoff J.M. Parker

Accurate sampling of the arterial input function (AIF) in low-temporal-resolution quantitative dynamic contrast-enhancedMRI (DCE-MRI) studies is crucial for accurate and reproducibleparameter estimation. However, when conventional AIFs aresampled at low temporal resolution, they introduce an unpre-dictable degree of error. An alternative double contrast agent(CA) bolus injection protocol designed to compensate for tem-poral mis-sampling of the AIF and tissue uptake curve wassimulated in addition to a commonly used single CA bolusinjection protocol. A range of tissue uptake curves for each AIFform were generated using a distributed parameter model, andMonte Carlo simulation studies were performed over a range ofoffset times (to mimic temporal mis-sampling), temporal reso-lutions and SNR in order to compare the performance of bothAIF forms in compartmental modeling. Insufficient data sam-pling of the single bolus AIF at temporal resolutions in excess of9 s leads to large errors, which can be reduced by employing anadditional, appropriately administered, second CA bolusinjection. Magn Reson Med 56:611–619, 2006. © 2006 Wiley-Liss, Inc.

Key words: dynamic contrast-enhanced MRI; quantitative anal-ysis; arterial input function; contrast agent; tumor

Quantitative dynamic contrast-enhanced (DCE) MRI pa-rameterizations of tissue microvasculature can be derivedthrough the use of tracer kinetic modeling techniques thatdescribe tissue permeability, perfusion, and the volume ofextravascular-extracellular space within the tissue of in-terest. Such techniques have important clinical value for awide range of pathologies, including malignancies (e.g.,Ref. 1 and references therein), multiple sclerosis (2,3),rheumatoid arthritis (4), atherosclerosis (5,6), Crohn’s dis-ease (7), and other inflammatory conditions (8).

The accuracy and precision of parameterizations gainedfrom compartmental modeling (9–11) and in particular theerrors introduced in the definition of the arterial inputfunction (AIF) has attracted considerable research interest(12–15). In theory, the use of an individually measuredAIF will increase the accuracy with which parameteriza-tions may be obtained relative to the use of an assumedAIF (16,17). It has also been suggested that measuring anAIF aides the reproducibility of parameter estimation byaccounting for day-to-day variation in contrast agent (CA)

administration and physiological factors that may affectCA circulation (17). However, measurement of an AIF isitself subject to errors, which can potentially lead to de-graded parameter accuracy and reproducibility. For exam-ple, errors are introduced if an AIF is sampled at a lowtemporal frequency, leading to it being mis-sampled rela-tive to its true shape (11) (Fig. 1). In particular, the firstpass of CA may be severely mis-sampled if a low-temporal-resolution dynamic acquisition protocol is used. This mis-sampling may be important, as much of the informationregarding capillary permeability and vascular volume,when using low-molecular-weight CAs such as Gd-DTPA,is highly dependent on accurate sampling of both the AIFand the tissue uptake curve during the first-pass period ofCA distribution.

The most common form of AIF used for DCE-MRI stud-ies is obtained using a single bolus CA injection. However,alternative injection protocols may be designed to reducethe impact of temporal mis-sampling on the measurementof the AIF bolus passage. We investigated the possibleadvantage of extending the number of peaks in the AIF(i.e., the number of CA bolus injections) from one to two.We hypothesized that in a scenario where the first peak inthe AIF (and in the tissue CA uptake curve) is mis-sampled(Fig. 1), a second bolus, if appropriately administered,could then compensate for a mis-sampled first bolus andsubsequently improve the overall accuracy and reproduc-ibility of the kinetic analysis.

MATERIALS AND METHODS

Data Simulation

Thirty-nine tissue uptake curves were simulated usingMatlab (Mathworks, Natick, MA, USA) and the adiabaticapproximation to the tissue homogeneity (AATH) model(18,19). This model is a distributed parameter model thatprovides a realistic approximation of in vivo data (9) andaccounts for blood plasma flow (Fp), permeability-surfacearea product (PS), and mean capillary transit time (Tc), inaddition to the extravascular-extracellular space (ve):

Ct�t� � Fp�0

Tc

Cp�t � t��dt�

� EFp�Tc

t

Cp�t � t��exp� � EFp�t� � Tc�

ve�dt� [1]

The volume transfer constant of CA between the bloodplasma and ve, Ktrans, is defined as the product EFp (20),

Imaging Science and Biomedical Engineering, University of Manchester,Manchester, UK.*Correspondence to: Caleb Roberts, Imaging Science and Biomedical Engi-neering, Stopford Building, University of Manchester, Oxford Road, Manches-ter, Greater Manchester, UK M13 9PT.E-mail: caleb.roberts@manchester.ac.ukReceived 31 January 2006; revised 5 April 2006; accepted 25 April 2006.DOI 10.1002/mrm.20971Published online 20 July 2006 in Wiley InterScience (www.interscience.wiley.com).

Magnetic Resonance in Medicine 56:611–619 (2006)

© 2006 Wiley-Liss, Inc. 611

where E is the extraction fraction of blood plasma from thecapillary, E � 1 � exp(�PS/Fp). For tissue time-coursesimulation, a range of values of E, Fp, and Tc were chosen,with ve kept at a steady value throughout (Table 1); thisparameter set is representative of data seen in the literature(21).

Figure 2 illustrates the process of data simulation. Signalintensity (SI) data were calculated for each simulated tis-sue and blood plasma concentration time course, based ona 3D spoiled gradient-echo (FLASH/fast field echo) DCE-MRI protocol (22) with tissue and blood T1 assumed to be1000 ms and 1400 ms, respectively (23). Each time courseconsisted of 1600 volumes covering 6 min with a temporalresolution of 0.23 s, a flip angle of 20°, and a TR of 2.5 ms.Two AIF injection protocols were simulated for subse-quent use as input to the data simulation process. Thesingle-bolus injection AIF (AIF1) was based on a popula-tion AIF obtained from 113 single AIFs calculated from 23patients enrolled on multi-visit DCE-MRI cancer trial (24)(Fig. 3). The simulated CA injection is based on0.1 mmol/kg Omniscan (Gd-BMA-DTPA, Gadodiamide)injected into the blood stream via the antecubital vein

using a power injector at a rate of 3 ml/s. A double-bolusinjection AIF (AIF2) was simulated using the same concen-tration and total volume of CA as used for AIF1, with theinjection rate for AIF2 halved relative to AIF1 to generatecomparable bolus widths (Fig. 3). For AIF2, the second CAbolus was administered with an offset of a 1⁄2 samplinginterval delay relative to the sampling of the first peak (Fig.3a). The first bolus of each injection protocol was designedto arrive in the tissue of interest 30 s after the onset of thedynamic series acquisition.

The SI, with baseline SI, S0, set arbitrarily to 100, wascalculated for tissue and blood time courses using thestandard relationship relating T1 to signal for a spoiledgradient echo (25) (assuming negligible T*2 signal loss) andassuming a CA relaxivity of 4.2 s–1mM–1.

Sample Offset and Addition of Noise

In total, four experiments were performed for each AIFform to investigate the effects of data mis-sampling. Priorto each experiment both the simulated tissue uptake curveand AIF were down-sampled from the initial high-tempo-

FIG. 1. Poorly sampled AIFs measured in vivo. AIFcalculated from the iliac artery with 2.3 s samplingtime (a), aorta with 4.5 s sampling time (b), andexternal carotid artery with 5.9 s sampling time (c).Mis-sampling, indicated by an arrow, is manifestedby a “shoulder” on the first-pass peak. The exam-ples were obtained from separate clinical studies(8,21,36).

Table 1Experimental Conditions (a) Showing Number of Permutations and Noise Levels Used for Each Experiment; Sampling Offset TimesUsed in the Data Simulation Process (b)*

Kinetic Parametersa SNR Iterations

E � 0.1, 0.5, 1.0Fp � 0.3, 0.5, 0.8 2, 3, 4.5, 7, 9, 100Ktrans � 0.018 to 0.48 13.5, 27 and �Tc � 0.01, 0.05, 0.1, 0.5, 1.0vp � 0.0018 to 0.18

(a) ve � 0.3

Simulated offset times for each sampling interval (s)Experiment 1 0.00 0.23 0.46 0.69 0.92 1.15 1.38 1.61 1.84 2.07

2 0.00 0.46 0.92 1.38 1.84 2.30 2.76 3.22 3.68 4.143 0.00 0.92 1.84 2.76 3.68 4.6 5.52 6.44 7.36 8.28

(b) 4 0.00 1.84 3.68 5.52 7.36 9.20 11.04 12.88 14.72 16.56

*Each time represents the offset of the simulated curves relative to the data sampling point. Data sampled at offset time � 0.00 s capturesbest representation of AIF and tissue uptake curve. All combinations of kinetic parameters were used to simulate data where FpTc, or vp,� 0.18 and EFp, or Ktrans, � 0.48 min�1—in total 39 combinations.aUnits: E (fraction), Fp (ml 100ml�1 min�1), Ktrans (min�1), Tc (min), vp (fraction), ve (fraction).

612 Roberts et al.

ral resolution tissue uptake and AIF curves to produce fourseparate data sets with time resolutions of 2.3 s, 4.6 s, 9.2 s,and 18.4 s. For each down-sampled data set, a total of 10further tissue uptake and AIF curves were generated at 10equally spaced offset times relative to each sampling in-terval to investigate the effect of data mis-sampling onparameter values (11) (Table 1). In addition to performingthe experiments without the addition of noise, zero meanGaussian noise, resulting in signal-to-noise ratios (SNRs)of 2, 3, 4.5, 7, 9, 13.5, and 27 were added prior to conver-sion to concentration-time curves. The SNRs were calcu-lated as the mean of the baseline SI signal divided by thestandard deviation (SD) of the noise.

Model Fitting and Statistics

We use the Kety tracer kinetic model extended to includea term for vascular plasma volume, vp (26) and employ anunconstrained nonlinear optimization fitting algorithm(Matlab, Mathworks, Natick, MA, USA) to examine theeffects of the different AIFs on model accuracy and preci-sion. This widely used compartmental model describesthe distribution of low-molecular-weight contrast media,such as Gd-DTPA, from the blood pool into the tissue

extracellular space at a rate determined by the blood flowto the tissue, permeability of the microvessel walls, andthe surface area of the perfusing vessels:

Ct�t� � vpCp�t� � Ktrans�0

t

Cp�t��exp�Ktrans�t � t��ve

�dt� [2]

For each time course in all experiments, 100 runs wereperformed to allow good sampling of the noise distribu-tions, and the median of each resultant model parameterwas calculated. The median parameter estimates at eachSNR were compared with the true simulation values, andthe absolute % error was calculated.

RESULTS

Examples of the two simulated noiseless AIFs used in theinitial data generation are shown in Fig. 3. These aresimilar in that AIF1 and AIF2 have the same peak widthbecause of the controlled CA injection procedure used.However, the first and second pass peaks of AIF2 arescaled down by a factor of 2 from AIF1 in order keep theCA volume identical for both injection protocols. Both AIFforms are used in the data simulation process to producetissue uptake curves with varying parameter characteris-tics (Fig. 4).

The shapes of the AIF and tissue uptake curves may becompromised due to mis-sampling of the data, particularlyat coarser temporal resolutions (Figs. 5 and 6). The mis-sampling of the first pass AIF peak results in AIF peak dataloss (the proportion of the first-pass peak not sampled) ofup to 16% at 2.3 s, 33% at 4.6 s, 66% at 9.2 s, and 83% ofthe AIF peak at 18.2 s temporal resolution. This is com-

FIG. 3. a: Schematic of double-bolus injection protocol. b: Simu-lated high-resolution single and double bolus injection AIFs thatwere used in the simulation process. [Color figure can be viewed inthe online issue, which is available at www.interscience.wiley.com.]

FIG. 2. Flow diagram showing the data simulation and analysisprocess. AIFPOP � population AIF obtained from (24). AATH �adiabatic approximation to the tissue homogeneity model, St � SItissue uptake curve.-

Single- and Multi-Bolus Contrast Injection 613

pensated for to a degree by the second, offset peak in AIF2

and the resulting tissue uptake curve, whereas with AIF1

this information cannot be retrieved.The peak sampling errors are dependent on the offset of

the peak(s) with respect to data sampling points. The de-gree of mis-sampling of first-pass data can determine theamount of error incurred in the estimation of both Ktrans

and vp (Fig. 7). As temporal resolution decreases, the errorassociated with mis-sampling is greater when a single bo-lus injection protocol is used (Fig. 8). As shown in Figs. 5and 6, at lower temporal resolutions, large portions of theAIF and tissue uptake curve are susceptible to mis-sam-pling, and in a worst-case scenario our simulations showthat vp estimates can have errors of more than 500%. Whenusing a double bolus injection protocol in combinationwith lower temporal resolution, these errors are reduced,even in worst-case scenarios, and show little variationwith sampling offset. In contrast to this, higher temporalresolution data are less susceptible to these errors, and theadvantage of AIF2 is negligible. Considering the dataacross all four temporal resolutions, sampling offsets, andtissue curves used in this study, half of the data (i.e., 25thto 75th percentile) for AIF1 have an error range of 5–88%for Ktrans and 40–500% for vp, whereas for AIF2 the errorrange is reduced at 6–42% for Ktrans and 42–240% for vp.

As expected, the performance of AIF1 and AIF2 worsenas SNR decreases (Fig. 9). In a scenario where no mis-sampling of the data occurs, the higher contrast-to-noiseratio (CNR) of AIF1 causes the parameterizations to be lesserror-prone in comparison with AIF2 (this is well demon-strated by Fig. 9c, in which the vp mean absolute error andrange are much lower for AIF1 across all SNR levels). SinceKtrans is determined more by the gradient of tissue contrastuptake rather than peak height (27), the error between bothAIF forms is similar. In a worst-case scenario and at lowtemporal resolution, AIF1 performs worse regardless ofSNR, with errors exceeding 50% for Ktrans and 500% forvp, even at high SNR. For experiments at high temporalresolution, however, the performance of AIF1 is superiorto AIF2 across all SNRs due simply to the minimization ofmis-sampling error (data not shown).

DISCUSSION

A precise and accurate assessment of the tumor microvas-cular environment relies on a number of important factorsduring the DCE-MRI examination. Assuming that accurateT1 values and subsequent CA concentration in the tissueand blood can be made, errors affecting the accuracy and

precision of DCE-MRI parameterizations can arise frominsufficient sampling of data, poor SNR, inadequate mod-eling techniques, and unregistered motion (10). In thisstudy we concentrated on the effect of CA administrationand insufficient sampling of the plasma and tissue timecourse.

In earlier DCE-MRI experiments (2,28) the AIF measure-ment, or calculation of plasma concentration (Cp), wastaken as an assumed function calculated from data with atemporal resolution of approximately 1 min. Henderson etal. (11) have since stated that the optimum sampling in-terval of the plasma time course should be on the order of1 s to capture the entirety of the first-pass peak, an AIFcharacteristic that is not accounted for in the assumed AIF.Investigators intending to conduct a DCE-MRI studyshould only set out to do so if they can optimize theirprotocol to acquire images as fast as possible (see below).

A single bolus contrast injection was previously shownto have advantages over an infusion injection (27) since abolus injection can reach a higher plasma concentration inless time than an infusion injection, which means thetissue will enhance more quickly with a smaller dose ofCA compared with the use of an infusion injection. Withcurrent MR protocols, DCE-MRI experiments that employhigh temporal resolution typically suffer from low SNRand spatial resolution, and therefore investigators may optto sacrifice high temporal resolution for a higher SNR andspatial resolution. In these cases an accurate and reliableAIF measurement is compromised and thus the wholesubsequent data analysis is subject to large errors, as wehave shown. The first-pass portion of the AIF is crucialbecause it provides important information regarding theinitial concentration of CA in the vascular space before itleaks into the extravascular-extracellular space at a ratedetermined by vessel wall permeability, and therefore it isthe clearest information available for estimating the vas-cular volume. In addition to insufficient sampling of theAIF, further complications include flow artifacts, partialvolume effects, the selection of the input voxels (espe-cially if the vessel of interest has any amount of motionthroughout the time course), and bolus dispersion anddelay (29). Since the AIFs used in this study were simu-lated, which excludes partial volume, input voxel selec-tion, and dispersion problems, the errors seen here can beattributed largely to insufficient sampling of the data andan inadequate modeling technique (we used a simplifiedmodel). It is also worthwhile to note that the rapid bolusAIFs used in this study are optimal for rapidly enhancing

FIG. 4. Example tissue uptake curves for AIF1 (a)and AIF2 (b). Each tissue uptake curve has zero-mean Gaussian noise added to create a SNR of 27in baseline SI (prior to conversion to CA concen-tration). Values of E, Fp, ve and Tc are 0.3, 0.5 mlml–1 min–1, 0.3 and 0.01 min (bottom curves), 0.5,0.5 ml ml–1 min–1, 0.3 and 0.1 min (middle curves),and 0.8, 0.5 ml ml–1 min–1, 0.3 and 1 min (topcurves), respectively.

614 Roberts et al.

FIG. 5. Examples of AIF2 at 2.3 s (a), 4.6 s (b), 9.2 s(c) and 18.4 s (d) sampling intervals to illustratehow error due to mis-sampling the AIF first-passpeak is compensated for by the second bolus AIFpeak (red) and vice versa (black). For each AIFexample the “true” AIF is shown as a dotted line toshow the degree of inaccuracies due to both mis-sampling the AIF peak and low temporal resolu-tion. Note the poor definition of the AIF as temporalsampling interval increases from 2.3 s to 18.4 s.

FIG. 6. Examples of tissue uptake curves gener-ated using AIF2 at 2.3 s (a), 4.6 s (b), 9.2 s (c), and18.4 s (d) sampling intervals to illustrate how errordue to mis-sampling the first-pass portion of thetissue uptake curve is compensated for by thesecond bolus peak (red) and vice versa (black). The“true” tissue uptake curve is shown as a dotted lineto show the degree of inaccuracies due to bothmis-sampling the AIF peak and low temporal fre-quency. Tissue parameters E � 0.3, Fp � 0.8 mlml–1 min–1, Tc � 0.1 min and ve � 0.3.

FIG. 7. Error in Ktrans (a) and vp (b) as a function of offset ofthe arrival of the first bolus peak as a fraction of the temporalresolution. Crosses (AIF1) and squares (AIF2) show the meanabsolute error, and bars show the range over all Ct curvesand all temporal resolutions investigated. SNR � 13 (typicalof clinical DCE-MRI studies). [Color figure can be viewed inthe online issue, which is available at www.interscience.wiley.com.]

FIG. 8. Error in Ktrans (a) and vp (b) as a function of temporalresolution over all sampling offsets. Crosses (AIF1) andsquares (AIF2) show the mean absolute error, and bars showthe range over all Ct curves and all sampling offsets investi-gated. SNR � 13 (typical of clinical DCE-MRI studies). [Colorfigure can be viewed in the online issue, which is available atwww.interscience.wiley.com.]

616 Roberts et al.

tissues, such as the liver and prostate, and may be lessnecessary for slowly enhancing tissues, such as the breast.

Errors induced by insufficient sampling of the AIF andtissue uptake curve are unknown prior to any DCE-MRIexperiment, since the offset of bolus arrival relative to thedata sampling is unknown. We used Monte Carlo simula-tions to provide an indication of the range of errors aninvestigator can expect given the temporal resolution andSNR of the dynamic data. Our simulations indicate theextent of error introduced by mis-sampling the first-passAIF peak and tissue uptake curve across a range of tempo-ral resolutions and SNR, and we introduced the idea ofutilizing a second contrast bolus to compensate for anymis-sampling that may occur in the first-pass of CA. Atrealistic levels of noise, and for data that are sampled atapproximately 9 s or longer, AIF2 (a double-bolus injectionAIF), with the second bolus delayed relative to the datasampling points by half a sampling interval (Fig. 3), islikely to improve the precision with which kinetic param-eters are evaluated. The tighter error range (Figs. 7–9) fromAIF2 indicates that it will perform better over a range ofSNR levels, tissue characteristics, and sampling offsets.The practicality of administering a double-bolus injectionis unlikely to cause any complication since a power injec-tor can easily be programmed to administer contrast twice,with each bolus separated by a defined period, especiallyat lower temporal resolution, where we have demonstratedthe largest likely benefit.

The SNR of the experiment is critical for the AIF calcu-lation, with higher SNR increasing the precision of the

calculation (30). Our experiments show that AIF1 is on thewhole less vulnerable to low SNR than AIF2 when nomis-sampling of the data occurs, particularly for vp. WhileAIF1 has the potential to perform badly at low temporalresolution, regardless of SNR, AIF2 remains relatively ro-bust regardless of SNR and sample offset. If the investiga-tor can minimize the affect of insufficient data sampling byincreasing temporal resolution at the expense of SNR andspatial resolution, then AIF1 is more beneficial than AIF2.While current technology requires this trade-off betweentemporal resolution and SNR, the steady development ofbetter coils, higher field strengths, and optimized acquisi-tion sequences in the future may improve the quality ofimages without compromising temporal resolution, and inthis case AIF1 would be preferred.

The marginally superior performance of AIF1 at 2.3 stemporal resolution indicates that this temporal resolution(in the case of our simulations, at least) approaches thatrequired to cause sampling errors to be negligible (11).This can be seen in Fig. 5a, where the AIF maintains amore substantial first-pass peak compared to data sampledat 4.6, 9.2, and particularly 18.4 s. Even by compensatingfor data loss by employing AIF2, errors are not reduced,and therefore a single bolus injection protocol is superiorin this case. In any tracer kinetic study, an important aimshould be to achieve reliable AIF sampling. However, inreality most studies are carried out at much lower tempo-ral resolution, indicating the potential benefits of using adouble (fractionated) AIF protocol such as we propose.

FIG. 9. Error in Ktrans (a and b) and vp (c and d) as a function of SNR over all tissue permutations and temporal resolutions at a samplingoffset of zero (a and c) and 1⁄2 sampling interval (b and d) to reflect best- and worst-case sampling scenarios, respectively. Crosses (AIF1)and squares (AIF2) show the mean absolute error, and bars show the range over all Ct curves and all four temporal resolution time datainvestigated. [Color figure can be viewed in the online issue, which is available at www.interscience.wiley.com.]

Single- and Multi-Bolus Contrast Injection 617

While there are clear advantages to using a fractionatedAIF, there are certain limitations that must be considered.The double bolus AIF depends on accurate timing of thesecond bolus relative to the first, and in circumstanceswhere there is a change in heart frequency and thus car-diac output, or if there is compression of the venous sys-tem between bolus injections, there is a potential for flawsto occur in the double bolus AIF, which may reduce itsimpact. It may also be argued that a double bolus AIFessentially prolongs the effective bolus duration by split-ting it into two halves, causing the AIF to be more suscep-tible to the effects of motion. A further difficulty may arisein qualitative visual interpretation of the tissue contrastuptake curves, since a double bolus injection protocol willalter the characteristic form of the tissue uptake curve.This may cause difficulties in the qualitative interpretationof heuristic parameters such as the wash-in gradient, timeto peak, and wash-out gradient. However, we have dem-onstrated that quantitative tracer kinetic model-basedanalysis can benefit from this alternative form of AIF.

This study also highlights the drawbacks of simple ki-netic modeling in DCE-MRI data analysis. Any kineticmodel is open to a number of errors since by definition amodel simply portrays an estimation of the true underly-ing physiology (9). However, it is possible to closely matchtrue physiology by increasing the complexity of the mod-eling process. It is often the case, though, that SNR con-straints and low temporal resolution prevent the use ofcomplex models, and in this circumstance a simplifiedmodel must be used with the understanding that its rep-resentation of tissue physiology will be somewhat poorer.As an example, our simulations show that in the absenceof noise, across all permutations the errors in Ktrans and vp

estimates (derived using an extended version of the Ketymodel) are at best 5% and 25%, respectively. These errorsare comparable to those reported elsewhere (9,11). Theminimum observed error for parameters in this study is aresult of fitting to data created by the St. Lawrence and Leemodel (19), which is a more complex and physiologicallysuitable model. The extended Kety model does not explic-itly provide separate estimates for E and F; instead itcombines them to give the EFp product, Ktrans. Further-more, the mean capillary transit time is considered to bezero in this model, and therefore estimates of vp (the prod-uct of Fp and Tc) will be flawed (9). We know from recentstudies that vp is a particularly unreproducible parameter,so it is not surprising that error values are particularly high(24,31). In addition to this, tissue uptake curves generatedwith values of Tc on the order of 1 s result in errors wellabove 500% for vp (data not shown). Error is further in-creased when the AIF is mis-sampled, and it is particularlystriking that when data are sampled at temporal resolu-tions greater than 9 s, which is typical of the temporalresolution currently employed in some studies (32–35),errors for both Ktrans and vp are greatly increased.

A poorly sampled AIF, as shown in our simulation stud-ies, can magnify errors, and in clinical phase I/II trialswhere DCE-MRI parameters are being employed as biomar-kers (or indicators) of disease status, it is imperative thaterrors be kept to a minimum. Data with a suitably highSNR are desirable; however, if an investigator compro-mises temporal resolution to achieve this, then any quan-

titative analysis will be at risk of experiencing large errorsdue to potential mis-sampling of the first pass of CA, inboth the blood plasma and the tissue of interest. We haveshown that at low temporal resolution these errors can beconstrained by using a double bolus injection AIF, withimportant implications for improved reproducibility andtherefore sensitivity to therapy.

CONCLUSIONS

For greater precision in kinetic parameter evaluation, it isimportant to sample a well-defined AIF. However, giventhat mis-sampling of data occurs in clinical studies, asdemonstrated in Fig. 1, and that we have no knowledge ofthe degree of bolus sampling offset relative to samplingpoints that may occur, our results suggest that at temporalresolutions in excess of 9 s a double bolus injection pro-tocol would be beneficial. At these longer sampling inter-vals, AIF2 successfully compensates for information lost inthe first pass of CA by having an offset second bolus,without an increase in the dose of CA.

ACKNOWLEDGMENTS

We thank Sue Cheung and Lucy Kershaw for their helpwith this study.

REFERENCES

1. Jackson A, Buckley D, Parker G, editors. Dynamic contrast-enhancedmagnetic resonance imaging in oncology. 1st ed. Berlin/Heidelberg/New York: Springer; 2005. 311 p.

2. Tofts PS, Kermode AG. Measurement of the blood-brain barrier perme-ability and leakage space using dynamic MR imaging. 1. Fundamentalconcepts. Magn Reson Med 1991;17:357–367.

3. Larsson HB, Stubgaard M, Frederiksen JL, Jensen M, Henriksen O,Paulson OB. Quantitation of blood-brain barrier defect by magneticresonance imaging and gadolinium-DTPA in patients with multiplesclerosis and brain tumors. Magn Reson Med 1990;16:117–131.

4. Palosaari K, Vuotila J, Takalo R, Jartti A, Niemela R, Haapea M, Soini I,Tervonen O, Hakala M. Contrast-enhanced dynamic and static MRIcorrelates with quantitative 99Tcm-labelled nanocolloid scintigraphy.Study of early rheumatoid arthritis patients. Rheumatology (Oxf) 2004;43:1364–1373.

5. Wasserman BA, Smith WI, Trout 3rd HH, Cannon 3rd RO, Balaban RS,Arai AE. Carotid artery atherosclerosis: in vivo morphologic character-ization with gadolinium-enhanced double-oblique MR imaging initialresults. Radiology 2002;223:566–573.

6. Yuan C, Kerwin WS, Ferguson MS, Polissar N, Zhang S, Cai J, Hat-sukami TS. Contrast-enhanced high resolution MRI for atheroscleroticcarotid artery tissue characterization. J Magn Reson Imaging 2002;15:62–67.

7. Pilleul F, Godefroy C, Yzebe-Beziat D, Dugougeat-Pilleul F, Lachaux A,Valette PJ. Magnetic resonance imaging in Crohn’s disease. Gastroen-terol Clin Biol 2005;29:803–808.

8. Roberts C, Rushton V, Jackson A, Parker G. The use of DCE-MRI in theassessment of lacrimal and salivary glands in Sjogren’s syndrome pa-tients. In: Proceedings of the 11th Annual Meeting of ISMRM, Toronto,Canada, 2003. p 766.

9. Buckley DL. Uncertainty in the analysis of tracer kinetics using dy-namic contrast-enhanced T1-weighted MRI. Magn Reson Med 2002;47:601–606.

10. Evelhoch JL. Key factors in the acquisition of contrast kinetic data foroncology. J Magn Reson Imaging 1999;10:254–259.

11. Henderson E, Rutt BK, Lee TY. Temporal sampling requirements forthe tracer kinetics modeling of breast disease. Magn Reson Imaging1998;16:1057–1073.

618 Roberts et al.

12. Fritz-Hansen T, Rostrup E, Larsson HB, Sondergaard L, Ring P, Hen-riksen O. Measurement of the arterial concentration of Gd-DTPA usingMRI: a step toward quantitative perfusion imaging. Magn Reson Med1996;36:225–231.

13. Parker GJ, Jackson A, Waterton J, Buckley DL. Automated arterial inputfunction extraction for T1-weighted DCE-MRI. In: Proceedings of the10th Annual Meeting of ISMRM, Toronto, Canada, 2003. p 1264.

14. Workie DW, Dardzinski BJ. Quantifying dynamic contrast-enhancedMRI of the knee in children with juvenile rheumatoid arthritis using anarterial input function (AIF) extracted from popliteal artery enhance-ment, and the effect of the choice of the AIF on the kinetic parameters.Magn Reson Med 2005;54:560–568.

15. Yang C, Karczmar GS, Medved M, Stadler WM. Estimating the arterialinput function using two reference tissues in dynamic contrast-en-hanced MRI studies: fundamental concepts and simulations. MagnReson Med 2004;52:1110–1117.

16. Port RE, Knopp MV, Brix G. Dynamic contrast-enhanced MRI usingGd-DTPA: interindividual variability of the arterial input function andconsequences for the assessment of kinetics in tumors. Magn ResonMed 2001;45:1030–1038.

17. Rijpkema M, Kaanders JH, Joosten FB, van der Kogel AJ, Heerschap A.Method for quantitative mapping of dynamic MRI contrast agent up-take in human tumors. J Magn Reson Imaging 2001;14:457–463.

18. St Lawrence KS, Lee TY. An adiabatic approximation to the tissuehomogeneity model for water exchange in the brain: II. Experimentalvalidation. J Cereb Blood Flow Metab 1998;18:1378–1385.

19. St Lawrence KS, Lee TY. An adiabatic approximation to the tissuehomogeneity model for water exchange in the brain: I. Theoreticalderivation. J Cereb Blood Flow Metab 1998;18:1365–1377.

20. Tofts PS, Brix G, Buckley DL, Evelhoch JL, Henderson E, Knopp MV,Larsson HB, Lee TY, Mayr NA, Parker GJ, Port RE, Taylor J, WeisskoffRM. Estimating kinetic parameters from dynamic contrast-enhancedT(1)-weighted MRI of a diffusable tracer: standardized quantities andsymbols. J Magn Reson Imaging 1999;10:223–232.

21. Buckley DL, Roberts C, Parker GJ, Logue JP, Hutchinson CE. Prostatecancer: evaluation of vascular characteristics with dynamic contrast-enhanced T1-weighted MR imaging—initial experience. Radiology2004;233:709–715.

22. Li KL, Zhu XP, Waterton J, Jackson A. Improved 3D quantitative map-ping of blood volume and endothelial permeability in brain tumors. JMagn Reson Imaging 2000;12:347–357.

23. Zhu XP, Li KL, Kamaly-Asl ID, Checkley DR, Tessier JJ, Waterton JC,Jackson A. Quantification of endothelial permeability, leakage space,and blood volume in brain tumors using combined T1 and T2* con-trast-enhanced dynamic MR imaging. J Magn Reson Imaging 2000;11:575–585.

24. Parker G, Macdonald A, Cheung S, Buckley D, Jackson A, Jayson G,Roberts C. An experimentally-derived functional form for a populationaveraged high temporal resolution arterial input function. In: Proceed-

ings of the 13th Annual Meeting of ISMRM, Miami Beach, FL, USA,2005. p 2100.

25. Haase A. Snapshot FLASH MRI. Applications to T1, T2, and chemical-shift imaging. Magn Reson Med 1990;13:77–89.

26. Tofts PS. Modeling tracer kinetics in dynamic Gd-DTPA MR imaging. JMagn Reson Imaging 1997;7:91–101.

27. Tofts PS, Berkowitz BA. Measurement of capillary permeability fromthe Gd enhancement curve: a comparison of bolus and constant infu-sion injection methods. Magn Reson Imaging 1994;12:81–91.

28. Weinmann HJ, Laniado M, Mutzel W. Pharmacokinetics of GdDTPA/dimeglumine after intravenous injection into healthy volunteers.Physiol Chem Phys Med NMR 1984;16:167–172.

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

30. Conturo TE, Akbudak E, Kotys MS, Chen ML, Chun SJ, Hsu RM,Sweeney CC, Markham J. Arterial input functions for dynamic suscep-tibility contrast MRI: requirements and signal options. J Magn ResonImaging 2005;22:697–703.

31. Roberts C, Issa B, Stone A, Jackson A, Waterton JC, Parker GJ. Compar-ative study into the robustness of compartmental modeling and model-free analysis in DCE-MRI studies. J Magn Reson Imaging 2006;23:554–563.

32. Doshi A, Thomasson D, Garci-Eulate R, Wedam S, Wise B, Gupta S, YaoJ, Steinberg S, Liewehr D, Choyke P, Swain S. Dynamic contrast en-hanced MRI for predicting clinical response to bevacizumab and che-motherapy in previously untreated inflammatory breast cancer: a com-parison of analytic methods. In: Proceedings of the 13th Annual Meet-ing of ISMRM, Miami, FL, USA, 2005. p 745.

33. Espinosa LA, Daniel BL, Vidarsson L, Zakhour M, Ikeda DM, HerfkensRJ. The lactating breast: contrast-enhanced MR imaging of normaltissue and cancer. Radiology 2005;237:429–436.

34. Liu G, Rugo HS, Wilding G, McShane TM, Evelhoch JL, Ng C, JacksonE, Kelcz F, Yeh BM, Lee Jr FT, Charnsangavej C, Park JW, Ashton EA,Steinfeldt HM, Pithavala YK, Reich SD, Herbst RS. Dynamic contrast-enhanced magnetic resonance imaging as a pharmacodynamic measureof response after acute dosing of AG-013736, an oral angiogenesisinhibitor, in patients with advanced solid tumors: results from a phaseI study. J Clin Oncol 2005;23:5464–5473.

35. Padhani AR, MacVicar AD, Gapinski CJ, Dearnaley DP, Parker GJ,Suckling J, Leach MO, Husband JE. Effects of androgen deprivation onprostatic morphology and vascular permeability evaluated with MRimaging. Radiology 2001;218:365–374.

36. Jayson GC, Parker GJ, Mullamitha S, Valle JW, Saunders M, BroughtonL, Lawrance J, Carrington B, Roberts C, Issa B, Buckley DL, Cheung S,Davies K, Watson Y, Zinkewich-Peotti K, Rolfe L, Jackson A. Blockadeof platelet-derived growth factor receptor-beta by CDP860, a human-ized, PEGylated di-Fab’, leads to fluid accumulation and is associatedwith increased tumor vascularized volume. J Clin Oncol 2005;23:973–981.

Single- and Multi-Bolus Contrast Injection 619