Intracranial tumor response to respiratory challenges at 3.0 T: Impact of different methods to quantify changes in the MR relaxation rate R2*

Authors


Abstract

Purpose:

To compare two ΔR2* quantification methods for analyzing the response of intracranial tumors to different breathing gases. The determination of changes in the magnetic resonance imaging (MRI) relaxation rate R2* (ΔR2*), induced by hyperoxic and hypercapnic respiratory challenges, enables the noninvasive assessment of blood oxygenation changes and vasoreactivity.

Materials and Methods:

Sixteen patients with various intracranial tumors were examined at 3.0 T. The response to respiratory challenges was registered using a dynamic multigradient-echo sequence with high temporal and spatial resolution. At each dynamic step, ΔR2* was derived in two different ways: 1) by subtraction of R2* values obtained from monoexponential decay functions, 2) by computing ΔR2* echo-wise from signal intensity ratios. The sensitivity for detection of responding voxels and the behavior of the “global” response were investigated.

Results:

Significantly more responding voxels (about 4%) were found for method (1). The “global” response was independent from the chosen quantification method but showed slightly larger changes (about 6%) when ΔR2* was derived from method (1).

Conclusion:

Similar results were observed for the two methods, with a slightly higher detection sensitivity of responding voxels when ΔR2* was obtained from monoexponential approximation. J. Magn. Reson. Imaging 2010;32:17–23. © 2010 Wiley-Liss, Inc.

BLOOD OXYGENATION LEVEL-DEPENDENT (BOLD) magnetic resonance imaging (MRI) may provide additional information beyond that obtained by other MR techniques (1). In BOLD MRI deoxyhemoglobin acts as an endogenous contrast agent. Due to its paramagnetic nature, deoxyhemoglobin shortens the MR time constant for the decay of transverse magnetization (apparent spin–spin relaxation time, T2*), which can be assessed by using gradient-echo sequences. Changes in the MR relaxation rate R2* (=1/T2*) are directly proportional to those in the deoxyhemoglobin (dHb) concentration (2, 3). The venous dHb concentration can be modulated by hyperoxic and hypercapnic respiratory challenges, which change the blood oxygenation saturation and the cerebral blood flow/volume. Measuring R2* changes during the inhalation of these gases provides information on blood and tissue oxygenation (2, 4–7) as well as on cerebrovascular reactivity (8–13) noninvasively. Such measurements can translate into pivotal tumor parameters such as vessel functionality, maturation, and growth (14, 15). With the increasing number of treatment procedures targeting tumor vascularity, such measurements have the potential to select suitable patients and to allow for a more refined assessment of treatment response (14, 15). The high clinical impact of this method is underlined by numerous animal studies (5, 7, 14–19) and by first studies of human tumors (20–24).

The response to respiratory challenges is commonly determined by measuring signal changes in T2*-weighted gradient-echo images acquired at a distinct echo time (10–15, 17, 20, 23–25). Although straightforward in application the analysis of changes in the BOLD signal has inherent limitations. First, it is unable to provide MR sequence- and system-independent quantitative data, which limits the reproducibility and comparability of the results. Second, it is nonspecific to R2* changes, because it is also affected by changes of the longitudinal relaxation rate R1 (=1/T1) and the inflow of unsaturated spins. Inflow effects heavily depend on the pulse sequence used (flip angle, repetition time or steady-state acquisition) and are a well-known problem in monitoring respiratory challenges (26) because inhalation of carbogen, O2, and CO2/air induces changes in cerebral blood flow (27). Third, the dynamic range of the measurement is limited by the choice of TE, which means that only a small range of tissue R2* values can be measured accurately.

The calculation of ΔR2* from multigradient-echo data overcomes these limitations (5, 9, 16, 21, 22). Two distinct approaches to derive ΔR2* from multigradient-echo data have been proposed:

  • 1ΔR2* is calculated by subtracting two R2* values, which are obtained from monoexponential approximation of the multiecho signal decay, eg, using exponential fit routines or other numerical algorithms (9, 16, 21). Those R2* results are independent from R1 changes and inflow effects. However, this technique is prone to errors since the relaxation behavior may deviate from an exponential function due to large-scale B0 field inhomogeneities (28, 29).
  • 2To become independent of the decay curve characteristic, ΔR2* can also be calculated echo-wise from ratios of signal intensities obtained from the T2*-weighted images. The inclusion of all echo-wise estimates can be achieved by appropriate weighted averaging (30). This approach requires the correction of artificial changes in the signal intensity caused by changes in R1 or inflow (apparent R1) (31).

To date, the theoretical advantages and limitations of both quantification methods for dynamic ΔR2* measurements during respiratory challenges have not been confirmed in a clinical setting. Thus, the purpose of this study was to compare both ΔR2* quantification methods in the assessment of intracranial tumors in response to hyperoxia and hypercapnia.

MATERIALS AND METHODS

Subject Population

Sixteen patients (eight men, eight women; age range, 30–74 years) with different intracranial lesions (six meningioma, three glioblastomas, five metastases, one non-Hodgkin-lymphoma, one tumor-like manifestation of clinical MS) were included. Only patients without signs of elevated intracranial pressure were included, as hypercapnia is known to further increase intracranial pressure. Lesion types were confirmed histologically in all cases, except for two patients with cerebral metastases of a known primary carcinoma and for one patient with the clinical diagnosis of multiple sclerosis (MS).

The Institutional Review Board approved this study. Written informed consent was obtained from all patients.

Gas Inhalation System

During the entire MR examination, patients were breathing spontaneously through a half-open breathing system consisting of a tightly fitting breathing mask, a two-way valve with a connection to the gas analysis unit, and two flexible inlet tubes for inspiratory and expiratory air stream. Without further connection to the inspiratory tube, patients were breathing room air. For the delivery of the breathing gases, the inspiratory tube was connected to a demand valve (Oxidem 3000, Dräger, Lübeck, Germany).

During repeated dynamic scans, each following a protocol of 1/4/2 minutes of air/gas/air inhalation, the following breathing gases (Linde Gas Therapeutics, Unterschleissheim, Germany) were applied: carbogen (Cb, 95% O2 + 5% CO2), CO2/air gas mixture (5% CO2 + 25% O2 + 70% N2), carbogen_light (Cb_light, 98% O2 + 2% CO2), and pure oxygen (100% O2). In 10 patients all four breathing gases, in five patients only Cb and CO2/air, and in one patient only Cb_light and O2 were investigated. The inspired oxygen concentration (FIO2) as well as the inspired and end-tidal CO2 partial pressures (PICO2 and PETCO2, respectively) were continuously monitored (Magnitude 3150/3155, Invivo, Orlando, FL).

MRI

MRI was performed at 3.0 T (Achieva, Philips Healthcare, Best, The Netherlands) using a transmit/receive quadrature headcoil. A dynamic 2D multigradient-echo multishot echo planar imaging (EPI) sequence with high temporal and spatial resolution (2.13 sec/frame, voxel size = 1.8 × 1.8 × 5 mm3) was used to measure a single transverse slice through the lesion center: field of view (FOV) = 230 × 201 mm2, 198 dynamic frames, 12 echoes (only odd-numbered echoes were acquired to achieve shortest possible minimum echo time but avoid inconsistencies due to systematic odd/even echo signal variations), TE = 4–83 msec, TR = 97 msec, flip angle = 25°, spectral presaturation by inversion recovery fat suppression, bandwidth in EPI frequency direction = 1290 Hz, EPI factor = 5. The chosen multishot variant allowed a smaller minimum TE value and a higher spatial resolution without decreasing the signal-to-noise ratio compared to a single-shot sequence. To achieve high B0 homogeneity, high-order volume shimming was used in all examinations. Total acquisition time of this sequence was 7 minutes.

MR Data Processing

The multigradient-echo data were processed offline by a custom-built research software (Philips Research Europe, Hamburg, Germany) based on an IDL 6.3 graphical user interface (Research Systems, Boulder, CO). All images were corrected for in-plane translational and rotational motion with respect to the data of the first dynamic frame, applying a rigid-body registration and correction algorithm (32). The transformation parameters were obtained from the registration of the first echo to be hereafter applied to all consecutive echoes of the dynamic frame.

At each dynamic step i, the MR signal was approximated by a monoexponential function of the echo-time TE, the relaxation rate R2*i and the signal intensity Mi at TE = 0:

equation image(1)

Changes of R2* at each dynamic step i, denoted ΔR2*i, were analyzed with respect to a reference dataset (index “ref”) defined as data average of the first three dynamic frames, using two different approaches.

Method A: ΔR2*i was expressed as the difference of two R2* values:

equation image(2)

R2*i was obtained from the monoexponential approximation of the multiecho signal decay (9, 16, 21) by integration of Eq. [1]:

equation image(3)

The Integral of Si(TE) was approximated numerically using the Simpson algorithm (33). R2*ref was also obtained in this way.

Method B: ΔR2*i was determined echo-wise by the logarithmic ratio of the signal intensities at dynamic step i and of the reference data, including all echo-wise calculations by weighted averaging (31) as described in the following. From Eq. [1], the change of R2* at dynamic step i can also be derived echo-wise according to:

equation image(4)

Differences in Mi and Mref may arise from signal intensity changes due to changes in R1 and inflow during the respiratory challenge. Therefore, Mi and Mref are extrapolated using the first three echoes of the signal decay [1]. The echo-wise estimates, given in Eq. [4], were combined by weighted averaging:

equation image(5)

δ(ΔR2*i(TE)) is the standard deviation of ΔR2*i(TE) and was approximated by the Gaussian error propagation function of equation [4].

Independent of the method used, the ΔR2*i time series were smoothed prior to response analysis by boxcar filtering with a width of three time frames. In order to identify voxels with significant R2* changes during a respiratory challenge, the following steps were performed: First, the intracranial region was manually segmented using the peak height image (sum of all multiecho images) of the first dynamic frame. Second, for each voxel a Student's t-test (P < 0.001, unequal variance) was used to compare the ΔR2*i values, acquired during the respiratory challenge, with those acquired during breathing room air (baseline data). For this comparison, a 60-second transitional period after changing the inhalation gas was excluded from analysis. Voxels with significant negative or positive R2* changes were displayed in a color-coded overlay map.

For quantitative response analysis, the voxel-wise ΔR2*i time series were averaged over all significantly changed voxels with an R2* decrease (expected predominant response type) to obtain a “global” response function ΔR2*i. After linear drift correction on account of the baseline data the “global” response functions were approximated by monoexponential functions:

equation image(6)

A nonlinear least-squares fit (Levenberg–Marquardt) to the experimental data yielded the following five parameters: A is the maximum change of the response function, τ1 and τ2 are the time constants of the rising and decaying exponential slopes, and t1 and t2 mark the start and endpoints of the response to the respiratory challenge. The fit quality is indicated by the coefficient of determination R2 (ie, 1 − [residual sum of squares χ2]/[data variance σ2]).

Statistical Analysis

The number of responding voxels within the whole brain as well as the behavior of the “global” response functions (expressed by maximum change A, the time constants τ1 and τ2 and the fit quality R2) were compared for the two ΔR2* quantification methods. Differences between the two methods were analyzed including all patients and all respiratory challenges using paired sample t-tests with a significance level of P = 0.05. All statistics were performed using SPSS 14.0 (Chicago, IL).

RESULTS

The procedure was well tolerated by all patients. PETCO2 changed significantly by +13(3) mmHg for Cb, by +15(2) mmHg for CO2/air, by +2(1) mmHg for Cb_light, and by −4(1) mmHg for O2 (P < 0.0001). FIO2 and PICO2 were 93(1)% and 37(1) mmHg for Cb, and 24(1)% and 38(2) mmHg for CO2/air, 96(2)% and 16(2) mmHg for Cb_light, and 97(2)% and 2(3) mmHg for O2, demonstrating the tightness of the breathing system.

The T2*-weighted images were of good quality with respect to susceptibility and motion artifacts in nine patients. In six patients, strong susceptibility artifacts arising from the frontal sinus were present in the images. The affected regions were excluded from the analysis.

In Fig. 1a an example of a patient with a meningioma is given. Strong susceptibility artifacts hampered the image quality frontally. In this case, a slightly larger amount of significantly changed voxels was found for Method B than for Method A (44.7% compared to 43.6% for carbogen, and 39.3% compared to 37.7% for CO2/air). In Fig. 1b an example of a patient with a metastasis is given. Again, strong susceptibility artifacts hampered the image quality frontally. In this case, a slightly smaller amount of significantly changed voxels was found for Method B compared to A (30.6% compared to 33.7% for carbogen_light and 21.6% compared to 21.9% for O2).

Figure 1.

Example of ΔR2* overlay maps obtained by the two methods for a patient (a) with a meningioma during Cb and CO2/air breathing and (b) with a metastasis during Cb_light and O2 breathing. Tumors are depicted on the T2-weighted reference images above. The overlays represent voxels with significant changes as a result of the performed t-test. Voxels with R2* decrease (ΔR2*<0) and increase (ΔR2*>0) are displayed color-coded with white to red and light blue to dark blue, respectively. In each map the color scales with the mean difference between the data obtained during the first and last minute of the dynamic scan (baseline) and during the last three minutes of the respiratory challenge, according to the data used in the t-test. The underlying images are peak height images (sum of all echoes) of the first dynamic frame. For each map the number of significantly changed voxels is given in percent of the total number of voxels within this brain area. When using Method B compared to A, in case (a) three maps showed an increased amount of responding voxels, one map showed a decreased amount, but the total amount of responding voxels was increased for both gases. In case (b) the amount of voxels with negative changes was decreased, whereas the amount of voxels with positive changes increased, but the total number of responding voxels was decreased for both gases.

Considering the whole patient group including all different respiratory challenges, it was found that in most cases (34 out of 52) the amount of significantly changed voxels was larger for Method A compared to Method B (up to 14%) and in 18 out of 52 cases the amount was smaller (up to 4%). The group analysis is given in Table 1. The total amount of responding voxels and also the amount of negatively changed voxels (ΔR2*<0) was significantly smaller for Method B compared to Method A (by 4% and 6%, respectively). The amount of positively changed voxels (ΔR2*>0) did not differ significantly between the two methods.

Table 1. Results for the Two ΔR2* Quantification Methods Including All Data (16 Patients)
N=52ΔR2*
Method AMethod BP
  1. For both ΔR2* quantification methods, the results of the parameterization of the “global” response functions belonging to the voxels with negative response as well as the results of the t-test are given as mean and standard deviation over the included data (all subjects and respiratory challenges).

  2. # denominates the total number of significantly changed voxels, #n and #p are the numbers of those with negative (ΔR2*<0) and positive (ΔR2*>0) response. They are all given in percent of total voxels within the segmented region.

A [s−1]−2.31 ± 0.83−2.17 ± 0.700.003
τ1 [s]51 ± 2847 ± 24n.s.
τ2 [s]26 ± 1228 ± 15n.s.
R20.858 ± 0.1860.870 ± 0.115n.s.
#n [%]34 ± 1232 ± 12<0.00001
#p [%]14 ± 714 ± 7n.s.
# [%]48 ± 1246 ± 12<0.0001

The “global” temporal response behavior (time constants τ1 and τ2) and the fit quality (R2) also did not differ significantly between the two methods. However, the maximum change A was found to be significantly smaller for Method B by about 6%. Also the interindividual variation of A was smaller for Method B (about 16%).

Moreover, it was shown that these findings are also valid when data were separately analyzed for images with and without macroscopic susceptibility artifacts from the frontal sinus (Tables 2, 3, respectively). However, differences in case of macroscopic susceptibility artifacts were of lower significance.

Table 2. Results for the Two ΔR2* Quantification Methods Including Only Data Related to Images With Strong Susceptibility Artifacts (6 Patients)
N=22ΔR2*
Method AMethod BP
  1. For further details see Table 1.

A [s−1]−2.29 ± 0.83−2.04 ± 0.540.02
τ1 [s]55 ± 3345 ± 21n.s.
τ2 [s]23 ± 727 ± 16n.s.
R20.839 ± 0.2480.876 ± 0.108n.s.
#n [%]31 ± 1329 ± 140.002
#p [%]15 ± 716 ± 80.021
# [%]46 ± 1145 ± 110.080
Table 3. Results for the Two ΔR2* Quantification Methods Including Only Data Related to Images Without Susceptibility Artifacts (10 Patients)
N=31ΔR2*
Method AMethod BP
  1. For further details see Table 1.

A [s−1]−2.30 ± 0.84−2.23 ± 0.790.018
τ1 [s]48 ± 2449 ± 25n.s.
τ2 [s]29 ± 1429 ± 14n.s.
R20.870 ± 0.1240.864 ± 0.121n.s.
#n [%]36 ± 1034 ± 100.0005
#p [%]12 ± 712 ± 7n.s.
# [%]48 ± 1347 ± 13<0.0001

DISCUSSION

This work was motivated by the fact that R2* values, obtained from monoexponential approximation of the multiecho signal decay, and also the ΔR2* results can strongly be hampered by large-scale magnetic B0 field inhomogeneities. Such B0 field inhomogeneities may arise from susceptibility gradients at tissue interfaces (sinuses, transition skull–air, auditory canal, etc.). Through-plane susceptibility effects deform the exponential decay of the multigradient-echo signal. They can be corrected (34), but this correction requires the measurement of neighboring slices, which is time-consuming and reduces the number of dynamic measurements of the slice of interest in a given time.

Recently, an alternative ΔR2* quantification method for dynamic multigradient-echo data was developed. This method calculates ΔR2* from signal intensity ratios including the multiecho data by weighted averaging (30). The used weighting function takes into account information about the noise in the T2*-weighted images and the relaxation times. However, this approach requires careful correction of artificial changes in the signal intensity (may be changed due to changes in R1 or inflow), which may be a limitation of the method (31). First applications of this method were performed in (U)SPIO-based blood volume and vessel size imaging (30, 35, 36) and in monitoring respiratory challenges (31, 37). In both applications a wide range of arbitrary ΔR2* values has to be analyzed and apparent R1 effects have to be correctly eliminated.

To date, the theoretical advantages and limitations have not been confirmed in a clinical setting. Thus, in this work we compared the two ΔR2* quantification methods for dynamic multigradient-echo data acquired during respiratory challenges in a group of patients with intracranial tumors. To our knowledge, this is the first formal evaluation of these ΔR2* quantification methods for dynamic multigradient-echo data of brain imaging and during respiratory challenges.

Comparing both ΔR2* quantification methods, it was found that:

  • 1The behavior of the “global” response function was independent from the used quantification method with the exception that the maximum change of the response was slightly larger when ΔR2* was determined by means of monoexponential approximation.
  • 2The total amount of responding voxels as well as the amount of voxels with negative response (ΔR2*<0) was found to be significantly larger when ΔR2* was calculated by means of monoexponential approximation regardless of the presence or absence of macroscopic susceptibility artifacts in the images.

Our study design accounts for the following aspects. First, in order to evaluate both methods over a large physiological range of data, we examined a group of patients with intracranial lesions. Thus, a wide spectrum of tissues with largely variable underlying R2*, ΔR2* and ΔR1 values and vascular properties (blood flow and volume) was investigated in different anatomic locations with variable macroscopic susceptibility. Moreover, the data acquisition was influenced by a likely lower compliance regarding patient motion than for example in volunteer groups. Second, in order to allow for a voxel-wise quantification with high precision and sensitivity, dynamic multigradient-echo images of a single slice were acquired with high temporal and spatial resolution. Third, in order to include data from all commonly used hyperoxic and hypercapnic respiratory challenges, we investigated the response to four different mixtures of O2 and CO2 breathing gases. Fourth, in order to reach sufficient statistical power of the analysis, we included all significantly changed voxels within a map, all patients, and all investigated respiratory challenges, even if the calculated mean values of the investigated parameters have large standard deviations and do not relate to a specific tissue type or to a defined state of hypercapnia or hyperoxia.

Even if we solely intended to evaluate the differences of two ΔR2* quantification methods with respect of the detection sensitivity and response behavior and not to characterize the tissue-specific behavior of various tumor entities to respiratory challenges, the following findings should be mentioned. Regardless of the quantification method, we found in our patient group a large amount of voxels with “inverse” response (ΔR2*>0) consistent with an increase in dHb content (2, 4). In healthy tissue of volunteers, for carbogen and CO2/air an R2* decrease was consistently detected (31) as a result of the increased oxygen saturation and cerebral blood flow. In contrast to carbogen and CO2/air, for pure oxygen the cerebral blood flow was found to be reduced (27) and a positive response may occur for voxels, where this effect exceeds the effect of improved oxygenation. In a tumor, the underlying mechanisms for R2* changes are complex and have been only investigated in animal studies (5, 15, 38–41). Under hypercapnia, areas with a predominantly positive response may be related to a vascular steal effect leading to a decreased blood flow or to the presence of immature vessels with minor vasoreactivity and blood volume increase as passive effect in case of increased blood pressure (7). Our findings of “inverse” response (ΔR2*>0) within tumors are consistent with observations of other groups (20, 21, 42). However, larger cohorts of patients are needed to fully characterize typical tumor behavior and to correlate these findings with histology. In this context, the application of multispin-echo sequences for R2 quantification is interesting, because that sequence type is only sensitive to vessels with a diameter less than 30 μm (4, 40, 43), which allows the specific assessment of the microvasculature. Spin-echo imaging would also eliminate the impact of large-scale field inhomogeneities; however, at the expense of sensitivity and temporal resolution. The use of parallel imaging techniques would improve temporal resolution, but the presently available phased array surface head coil did not provide enough free inner space to apply the face mask for gas breathing.

In the present application we found similar results for the two methods, with slightly higher detection sensitivity and maximum change of the “global” response, if ΔR2* was calculated from monoexponential approximation of R2*. The differences between the two methods were not larger in cases with macroscopic susceptibility artifacts compared to those without artifacts. As we cannot provide a gold standard of voxel response to decide which method is more accurate, the overall detection sensitivity of the response to respiratory challenge appears to be the most favorable argument for one method over the other.

In conclusion, both ΔR*2 quantifications, from signal intensity ratios as well as from monoexponential approximation, are suited to monitor R2* changes in patients with intracranial tumors during respiratory challenges. However, using the monoexponential approximation method a slightly higher detection sensitivity of responding voxels can be achieved. Thus, this work may be helpful in deciding which method and processing strategies are to be used in this upcoming field of research.

Ancillary