- Email: [email protected]
Measurement of regional brain temperature using proton spectroscopic imaging: validation and application to acute ischemic stroke
Magnetic Resonance Imaging 24 (2006) 699 – 706
Measurement of regional brain temperature using proton spectroscopic imaging: validation and application to acute ischemic stroke Ian Marshall4, Bartosz Karaszewski1, Joanna M. Wardlaw, Vera Cvoro, Karolina Wartolowska2, Paul A. Armitage, Trevor Carpenter, Mark E. Bastin, Andrew Farrall, Kristin Haga SHEFC Brain Imaging Research Centre for Scotland, University of Edinburgh, Edinburgh, UK Received 15 October 2005; accepted 5 February 2006
Abstract A magnetic resonance proton spectroscopic imaging (SI) technique was developed to measure regional brain temperatures in human subjects. The technique was validated in a homogeneous phantom and in four healthy volunteers. Simulations and calculations determined the theoretical measurement precision as approximately F0.38C for individual 1-ml voxels. In healthy volunteers, repeated measurements on individual voxels had an S.D. = 1.28C. In a clinical study, 40 patients with acute ischemic stroke were imaged within 26 h (mean, 10 h) of onset. Temperatures were highest in the region that appeared abnormal (i.e., ischemic) on diffusion-weighted imaging (DWI) compared with a normal-appearing brain. The mean temperature difference between the DWI blesionQ area and the b normal brain Q was 0.178C [ P b 103; range, 2.458C (hotter) – 2.178C (cooler)]. Noninvasive temperature measurement by SI has sufficient precision to be used in studies of pathophysiology in stroke and in other brain disorders and to monitor therapies. D 2006 Elsevier Inc. All rights reserved. Keywords: Spectroscopic imaging; Temperature; Brain; Stroke
1. Introduction Acute brain injury from a number of causes (e.g., stroke) is common and is associated with elevated body temperature [1–3], which in turn is associated with poor functional outcome. The cause of elevated body temperature and its relationship to brain temperature are unknown. A more severe stroke is associated with higher body temperatures [1,4]. In experimental models, cooling (hypothermia) may reduce lesion size [1,5] and improve outcome [1,6,7]. There is evidence that the same may be true clinically [8]. Direct
This work was presented, in part, at the 22nd Annual Meeting of ESMRMB, Basel, Switzerland, September 2005 (abstract 320). 4 Corresponding author. Medical Physics, Western General Hospital, EH4 2XU Edinburgh, UK. Tel.: +44 131 537 1661; fax: +44 131 537 1026. E-mail address: [email protected] (I. Marshall). 1 Current address: Department of Neurology of Adults, Medical University of Gdansk, Gdansk, Poland. 2 Current address: Department of Human Anatomy and Genetics, University of Oxford, Oxford, UK. 0730-725X/$ – see front matter D 2006 Elsevier Inc. All rights reserved. doi:10.1016/j.mri.2006.02.002
measurement of brain temperature might provide insights into the pathological processes taking place in ischemic and infarcted tissues [9]. Conventional methods of measuring brain temperature require invasive probes and can monitor only one or a few locations simultaneously [9,10]. Clearly, a reliable noninvasive method is highly desirable and could have many clinical uses. Several magnetic resonance parameters can be used for the noninvasive measurement of regional temperatures [11], including the diffusion coefficient, the longitudinal relaxation time constant (T 1) and the proton resonance frequency (PRF). Methods based on water PRF, which has a temperature dependence of approximately 0.01 ppm/8C [11], are the most popular. In bphase-shiftQ PRF techniques, the pixel-by-pixel phase change between successive gradient-echo images is used to monitor thermal interventions such as forced heating [11,12]. Although rapid and yielding spatial resolutions of imaging, these techniques are restricted to monitoring temperature changes and, thus, are not suitable for observational studies. The second class of PRF techniques uses spectroscopic acquisition to measure water
700
I. Marshall et al. / Magnetic Resonance Imaging 24 (2006) 699 – 706
chemical shift relative to a temperature-independent reference chemical shift [13 –15]. The reference compound depends on the application [e.g., lipids have been used for liver imaging [15], while N-acetyl (NA) compounds are most suitable for brain imaging [13,14]. These spectroscopic techniques are, in principle, capable of estimating absolute temperatures. Localized brain temperatures in experimental animals [14,16], healthy adults [17,18] and newborn infants [13] have been measured using single-voxel spectroscopy. However, the potential of spectroscopic imaging (SI) to map regional brain temperatures in diseases has not yet been exploited. Studies of brain temperature in stroke patients have so far been limited to single-point measurements using invasive sensors [9,10]. There have been no reports of brain temperature measurements made with magnetic resonance spectroscopic techniques in patients with acute ischemic stroke. Indeed, a theoretical problem for these techniques in ischemic stroke may be that the associated reduction in NA due to neuronal death may make it impossible to assess the relative water chemical shift. In this work, we validated a method of temperature measurement based on standard SI sequence in a test phantom and in four normal volunteers, and we used it in a study of stroke patients to measure temperature in an ischemic lesion [as defined on diffusion-weighted imaging (DWI)] and in a normal brain.
2. Methods 2.1. Theory of spectroscopic temperature measurement Temperature T is derived from the chemical shift of water (CSH2O) using a relation of the form: T ¼ Tref þ k ðCSH2 O CSref Þ
ð1Þ
where CSref is the (temperature-independent) chemical shift of a reference compound, k is the coefficient of proportionality and Tref is the reference temperature. In this work, scanner frequency was adjusted to bring the water signal onto resonance, and we constantly assigned a nominal value of 4.70 ppm to water frequency. The chemical shift of brain metabolites therefore appeared to vary with temperature, and a working version of Eq. (1) was given by: T ¼ 37 þ 100ðCSNA 2:035Þ
ð2Þ
where CSNA is the apparent chemical shift of NA compounds (expressed in ppm), the dominant singlet peak chosen for temperature determination in this work. A reference value of CSNA = 2.035 ppm was determined from occipital voxels in a group of 20 healthy control subjects (age, 46F10 years) with an assumed brain temperature of 378C. The constant of proportionality (1008C/ppm or, equivalently, 0.01 ppm/8C) was taken from the literature [11]. Eq. (2) was used in the simulation of
spectroscopic data and in the calculation of voxel temperatures from actual data. 2.2. Simulations Synthetic free induction decay (FID) signals were created to represent signals typically obtained during in vivo spectroscopic brain examinations on our 1.5-T scanner. FIDs consisted of 512 complex data points with a sample spacing of 1 ms and comprised a sum of exponentially decaying sinusoids corresponding to the major metabolites choline, creatine and NA compounds. Time domain amplitudes were 60, 60 and 100 U for the choline, creatine and NA components, respectively, and a common T 2 value of 300 ms was used. A bresidual water Q signal with an amplitude of 5000 U was added to represent the typical performance of our chemical-shift-selective (CHESS) water suppression across SI slices. FIDs were multiplied by a Gaussian envelope with a full width at half-maximum (FWHM) of 4 Hz to simulate line broadening due to intravoxel inhomogeneities [19]. In a first set of simulations (b temperature variation Q), the chemical shift of the metabolites (relative to water) was varied according to Eq. (2) to represent temperatures ranging from 368C to 398C in steps of 0.58C. At each chemical shift, a set of 100 FIDs was generated with Gaussian (i.e., normally distributed) noise [20] having an S.D. =50 U. This noise level is representative of 1-ml voxels on our scanner. In a second group of simulations (b NA amplitudeQ), the amplitude of the NA signal was varied from 100 to 20 U in steps of 10, with all other parameters being held constant. In a third set of simulations (bline broadening Q), the Gaussian line broadening was 4, 6, 8 or 10 Hz. Finally, the effect of noise level was investigated by generating a set of FIDs with a noise S.D. =20 U, with all other parameters being the same as for 378C FIDs in the first set of simulations. Groups of FID were processed in a standard manner. Briefly, the residual water component was first removed from each FID using the Hankel Lanczos Singular Value Decomposition technique [21]. Resulting signals were then quantified in the time domain using the AMARES algorithm [22] within the MRUI package [23]. A model consisting of three in-phase Gaussian components was used, as Gaussian lineshapes better describe in vivo data than Lorentzian lineshapes [19]. The mean and standard deviation of the NA peak chemical shift were determined for each group of FIDs and converted to equivalent temperatures using Eq. (2). 2.3. In vitro validation Initial in vitro experiments were carried out to determine measurement precision (standard deviation across the SI slice) under ideal conditions. A homogeneous phantom containing metabolites at physiological concentrations was imaged once a month as part of a quality assurance program. The phantom was kept in a scanner room to
I. Marshall et al. / Magnetic Resonance Imaging 24 (2006) 699 – 706
701
2.4. In vivo validation All human scannings were approved by the Local Research Ethics Committee. Initial in vivo evaluation was carried out on four healthy male volunteers (ages, 23– 38 years), who each had four SI scans within one examination. The SI slice was positioned at the level of the basal ganglia, and sequence parameters were the same as for the in vitro study. Volunteers were removed from the scanner and repositioned between the second and third scans. Spectroscopic data were processed as described above, and cerebral temperature for each valid voxel was calculated using Eq. (2). Statistical analysis was carried out using an autoregressive mixed model in SAS Version 9.1 (www.sas.com). 2.5. Stroke patients
Fig. 1. Determination of temperature from simulated FIDs representing in vivo signals from 1-ml voxels in the human brain at various temperatures. At each temperature, the mean and standard deviation are shown for a group of 100 FIDs with a Gaussian noise S.D. of 50 U, an NA level of 100 U and a Gaussian line broadening of 4 Hz. Actual simulated temperatures are shown by the line of identity (dashed line).
ensure thermal equilibrium, and its temperature was determined by a liquid crystal display (LCD) thermometer strip permanently attached to it. The reading accuracy was estimated as F18C. All magnetic resonance imaging (MRI) measurements were made on a 1.5-T Signa scanner (GE Healthcare, Slough, UK) fitted with standard head coil. The manufacturer’s single-slice SI sequence was used with PRESS excitation, an echo time (T E) of 145 ms and a repetition time (T R) of 1000 ms. The field of view (FOV) was 32 cm, with a phase-encoding grid of 2424 and a slice thickness of 10 mm. Automatic shimming and CHESS water suppression (with three pulses) were applied. For each phase encoding, 512 complex data points were acquired with a sampling interval of 1 ms. SI scanning time was approximately 10 min. Raw spectroscopic data were transferred to a workstation and interpolated to a 3232 matrix during Fourier transformation, thereby yielding voxels of nominal dimensions of 101010 mm. Zero-order phase correction was carried out on a voxel-by-voxel basis using residual water signals [24], thereby effectively bringing water to a fixed nominal chemical shift of 4.70 ppm. Subsequent processing and AMARES quantification were as described above. Spectra were automatically discarded if fitted linewidths were b1 Hz or N 10 Hz, or if metabolite peaks were N0.1 ppm offset of their expected values, or if voxels lay on the edges of the PRESS excitation region. All spectra were also inspected visually and discarded if judged to be of poor quality (e.g., having a badly elevated baseline or containing spurious peaks). Temperatures for each remaining voxel were calculated using Eq. (2).
We recruited 40 patients (mean age, 78 years; range, 58– 95 years) with symptoms of moderate to severe cortical ischemic stroke in an MRI study to compare metabolites in terms of their appearance on DWI. The study was approved by the Local Research Ethics Committee, and all subjects gave written informed consent. Patients had magnetic resonance brain examination within 26 h (mean, 10 h) of stroke onset. The examination included axial diffusionweighted spin-echo echo-planar imaging (DWI) and SI centered on the slice showing the maximum ischemic lesion extent on DWI. The imaging parameters for DWI were: FOV=240 mm, axial slice= 15, thickness = 5 mm, slice gap =1 mm, acquisition matrix = 128128 and T R/ T E = 10 s/97 ms. Diffusion-sensitizing gradients with scalar b values of 1000 s/mm2 were applied in six noncollinear directions. The total examination time, including patient setup, was typically 40 –50 min. SI parameters and data
Fig. 2. Determination of temperature from simulated FIDs representing in vivo signals from 1-ml voxels in the human brain with various levels of NA compounds. At each level, the mean and standard deviation are shown for a group of 100 FIDs with Gaussian noise S.D. of 50 U and a Gaussian line broadening of 4 Hz. The actual simulated temperature is 37.08C (dashed line).
702
I. Marshall et al. / Magnetic Resonance Imaging 24 (2006) 699 – 706
Fig. 3. (A) Mean and (B) standard deviation temperature maps of four successive measurements made at the level of the basal ganglia in a healthy volunteer. The temperature scale is given in degrees Celsius.
processing were as described above, except that a fivecomponent model was used for AMARES quantification. Prior knowledge was applied to link the two additional Table 1 Effect of Gaussian line broadening on temperatures determined from simulated FIDs representing in vivo signals from 1-ml voxels in the human brain Applied line broadening (Hz)
Temperature (meanFS.D.)
Fitted FWHM (Hz)
4 6 8 10
37.02F0.29 37.01F0.48 36.99F0.67 36.97F0.87
4.7 6.7 8.6 10.5
For each line broadening, the mean and standard deviation are shown for a group of 100 FIDs with a Brownian noise S.D. = 50 U, NA amplitude = 100 U, intrinsic T 2 = 300 ms and simulated temperature = 37.08C. Also shown is the Gaussian linewidth (FWHM) fitted by AMARES quantification.
Fig. 4. (A) Diffusion-weighted image and (B) corresponding temperature map of a patient with an acute ischemic stroke lesion in the left parietal region. In (B), the lesion outline is shown in red. The temperature scale is given in degrees Celsius.
components in order to model the doublet signal of lactate. Cerebral temperature for each valid voxel was calculated using Eq. (2). The SI voxel grid was superimposed on the DWI image. SI voxels overlying the DWI lesion were identified as b LES.Q Voxels containing a normal-appearing brain located beyond the DWI lesion were identified as b ipsilateral normal Q (INL) or b contralateral normal Q (CNL) in the ipsilateral and contralateral hemispheres, respectively. Voxels covering the cerebrospinal fluid were excluded. For each patient, mean temperatures were calculated for each tissue category, and mean temperature differences between categories were calculated. Finally, the group means of these values were calculated across all patients. Statistical analysis was carried out using odds ratios by fixed-effects
I. Marshall et al. / Magnetic Resonance Imaging 24 (2006) 699 – 706
703
model in RevMan software (http://www.cc-ims.net/ RevMan). This analysis was primarily within patients [i.e., comparing abnormal vs. normal temperature within each patient, thereby controlling for external factors (such as variation in scanner performance) that might vary between patients].
3. Results 3.1. Simulations In the first set of simulations, temperatures deduced from the quantification of synthetic FIDs were in excellent agreement with actual temperatures (Fig. 1). At each temperature, there was no bias in mean results, and the standard deviation of results was 0.38C. In the second group of simulations, the mean temperature was accurately determined at each level of NA (Fig. 2). The standard deviation increased in inverse proportion to the level of NA, as NA fell from 100 to 40 U. Below 40 U, the standard deviation increased more rapidly than in the inverse proportion. At an NA level of 20 U, the fitting algorithm failed in 5 of 100 cases, and the standard deviation of the remaining 95 cases was 2.58C, compared with a value of 1.58C extrapolated from higher levels and assuming an inverse linear relationship. The quality of these spectra was such that they would have been rejected in an experimental situation. The effect of Gaussian line broadening is shown in Table 1. The temperature was accurately determined for all line broadenings studied, while the standard deviation increased with increasing broadening. Reducing the noise standard deviation from 50 to 20 U reduced the standard deviation of temperature estimations from 0.308C to 0.128C. 3.2. In vitro validation Each set of voxel temperatures estimated from monthly SI measurements of the phantom had an S.D. b 0.28C. Over the observed seasonal range of 19–248C, the difference (meanFS.D.) between derived temperatures and LCD thermometer temperatures was 0.4F0.78C, which is within the reading accuracy of the LCD thermometer.
Table 2 Summary of regional brain temperature differences in a study of 40 patients with acute ischemic stroke Groups
Temperature difference (95% confidence interval) (8C)
n
LES versus INL+CNL LES versus INL LES versus CNL CNL versus INL
0.17 0.50 0.38 0.22
40 24 29 35
(0.07, (0.34, (0.27, (0.12,
0.27) 0.66) 0.49) 0.32)
All findings are statistically significant ( P b 103). n, number of contributing patients (see also Fig. 5).
Fig. 5. Histograms of intraindividual temperature differences (8C) between LES, INL and CNL brain tissues in a study of 40 patients with acute ischemic stroke (see also Table 2).
3.3. In vivo validation Useful data were obtained from approximately 30 voxels (mean = 29, S.D. =6) in each cerebral hemisphere of each volunteer on each scan. Spectral quality was low in the frontal region. The overall mean brain temperature (all subjects, all scans) was 36.58C. There was no significant difference between hemispheres. There was a small but significant ( P b.001) effect of scan number, with the temperature falling by 0.098C between successive scans. However, this effect was not significant for Subject 1 alone; therefore, we combined separate scan data for this subject for the purposes of Fig. 3. The statistical model gave a residual variance of 1.48C (i.e., repeated measurements on individual voxels had an S.D. = 1.28C). 3.4. Stroke patients Most spectra were of sufficient quality, and most lesions were large enough for each patient to yield temperature readings from at least two voxels (and, in some cases, more than 40 voxels) in each of the tissue groups (LES, INL and CNL). A typical example is shown in Fig. 4. Although ischemic stroke reduces NA levels, spectral peak areas were still sufficient to allow valid temperature measurements from many voxels in all patients. The group mean temperatures of different tissue classes were not significantly
704
I. Marshall et al. / Magnetic Resonance Imaging 24 (2006) 699 – 706
different. However, the group means of intraindividual temperature differences were highly significant ( P b103) (Table 2). Histograms of the differences are shown in Fig. 5. Lesion temperature was 0.178C higher than normal tissue (INL+CNL combined) temperature [range, 2.458C (hotter)–2.178C (cooler)], and CNL tissue temperature was 0.228C higher than INL tissue temperature [range, 1.088C (hotter)–1.568C (cooler)]. 4. Discussion Temperatures were estimated accurately for all simulated FIDs over the range of temperatures, NA levels and linewidths expected in vivo. Under ideal conditions (homogeneous spectroscopic phantom at constant uniform temperature), the SI technique yielded voxel temperatures with an S.D. = 0.28C across the measurement slice. Temperature determination using proton spectroscopic techniques requires estimation of the frequency of water and reference peaks. Given that the scanner will adjust to water resonance frequency, the problem reduces to that of determining the frequency of the reference peak. We used the singlet peak of NA for this work. Simulations suggested a frequency precision of 0.003 ppm and, thus, a temperature precision of 0.38C for typical 1-ml voxels in healthy volunteers. These simulations assumed a 4-Hz Gaussian line broadening of an NA signal having an intrinsic T 2 =300 ms and an amplitude =100 U, corrupted with noise S.D. = 50 U. A framework for the estimation of model parameters is provided by the Cramer–Rao (CR) theory [25,26]. The theory is based on the sensitivity of a model function to constituent parameters and can be used to evaluate the achievable precision of parameter estimates given the constraints of data and assuming bperfect Q quantification. The resulting CR boundary (CRB) is the best that can be achieved. An analytical expression of CRB on the frequency determination for a single Lorentzian peak has been given [27] as: pffiffiffi pffiffiffi r CRBN ¼ 2 2ð aÞ3=2 ts ð3Þ c where a is damping, t s is sampling interval, r is noise standard deviation and c is signal amplitude. All quantities relate to the time domain. Note that the ratio in parentheses is the inverse of the signal-to-noise ratio (SNR). For a Lorentzian lineshape, damping is related to FWHM by a =p FWHM. Evaluating Eq. (3) using the NA and noise parameters given above, together with a sampling interval of 1 ms and a FWHM of 4 Hz, gives CRBN =2.0 rad or 0.3 Hz. At a field strength of 1.5 T, this is equivalent to 0.005 ppm or 0.58C. This result is in reasonable agreement with the value of 0.38C found from simulations. The discrepancy is due to the fact that Eq. (3) holds strictly only for Lorentzian lineshapes (i.e., exponentially decaying FIDs) and will not be accurate for Gaussian and mixed lineshapes used in simulations and found in in vivo data
[19]. Corresponding expressions for lineshapes other than Lorentzian are not available, but Lorentzian calculation is offered here as a guide. Intuitively, we might expect a b worseQ result for Lorentzians than Gaussians, as a significant part of the energy of Lorentzian lineshapes is contained in extensive tails that are difficult to detect above the noise level. Eq. (3) shows that CRBN increases more than linearly with linewidth, and, thus, shimming is important in achieving precise temperature estimation. Temperature standard deviations in Table 1 increased with linewidth to a power of approximately 1.3. A temperature resolution of 0.38C requires a 0.003-ppm frequency resolution, which is 0.2 Hz at 1.5 T. To measure such a small difference in a conventional manner from a spectrum (i.e., a Fourier-transformed FID) would require an FID duration of 5 s, or 2.5 s if zero filling were applied. Given that intrinsic T 2 decay and T 2* processes limit the usable signal duration to a few hundred milliseconds at best, this is clearly not practical. A more sophisticated method is required for frequency determination from a limited duration signal. In this work, we used AMARES for the time domain modeling of complex FIDs. Effectively, the change in phase between successive data points is used to estimate the corresponding frequency. Modeling in the frequency domain is also possible [28], and both methods are, in principle, capable of giving results limited only by CR bounds. Temperature determination using proton spectroscopic techniques requires that both water peak and internal reference peak be visible in each spectrum. Spectroscopic data are normally acquired with water suppression so as not to swamp weak metabolite signals with overwhelming signal due to water. In clinical SI, there is insufficient time to also collect a nonsuppressed data set and, therefore, residual water signal must be large enough for frequency (hence, temperature) determination. In the standard implementation of SI that we used, we found that even with water suppression, the residual water signal was typically 30–50 times greater than metabolite signals. We did not need to deliberately b degrade Q water suppression. Of course, the selective RF pulse(s) used for water suppression should not distort the shape of the (residual) water peak; otherwise, an artifactual b shift, Q which is misinterpreted as a temperature difference, may be created. In the worst case, this effect could vary across the FOV if field inhomogeneities shifted the primary water resonance toward either edge of the frequency band of the pulse(s). In phantom experiments comparing suppressed and nonsuppressed water acquisition (data not shown), we were unable to detect any systematic differences in estimated temperatures. We are therefore satisfied that our standard water suppression technique is valid. We used the residual water signal for eddy current and zero-order phase correction [24]. We used the NA peak as the temperature-independent reference. NA is the dominant singlet peak in proton brain spectra acquired over a wide range of echo times and is, therefore,
I. Marshall et al. / Magnetic Resonance Imaging 24 (2006) 699 – 706
the obvious candidate. Even within DWI-infarcted (LES) tissue, we found that NA levels were typically 66% [29] and usually no lower than 30% of the level found in healthy brain tissue. This finding could be explained by partial volume effect, or by the fact that we imaged patients sufficiently soon after stroke while metabolite changes were still evolving. Whatever the explanation, we did not need to discard voxels solely because of insufficient NA signal. The slice selection gradient strength in our SI sequence was 6 mT/m for the 10-mm slice thickness used. With a nominal chemical shift difference of 2.7 ppm between water and NA, this leads to a slice position offset of 0.7 mm. We consider this theoretical offset to be negligible, as it is well within the uncertainty in slice prescription and subject head motion during a typical examination. The simulations and CR calculations suggest a theoretical temperature precision of 0.38C for individual wellshimmed 1-ml voxels in healthy volunteers. The residual S.D. of 1.28C for repeated measurements on individual voxels in four healthy volunteers was higher than what might be suggested by theoretical precision. Increased random variation may be due to subject motion, spectral contamination from lipids or scanner instabilities, and sets a realistic bound on what can presently be achieved in vivo. We hypothesize that the fall in temperature by 0.098C between successive scans was due to the combined effects of subject inactivity and cooling by the air flow in the scanner bore. Measurement repeatability is likely to be worse in patients when voxels with NA levels down to approximately 30% of normal levels might be accepted as valid. Thus, intraindividual temperature differences from several subjects need to be pooled in order to detect differences on the order of 0.18C, as reported here. It is too early to say whether the method can measure temperature in individuals. However, it is possible that it could, as temperature measurements are available in absolute units. To overcome concerns regarding precision and repeatability and the problem of different numbers of voxels contributing to the analysis in each patient, we have concentrated our analysis on within-patient differences so that each subject’s normal tissue acts as a control. This approach minimizes the effects of between-session variability (as yet not fully quantified for this technique) and has allowed us to obtain highly significant P values despite small differences in temperature. To further err on the side of caution, we have used grouped bwithin-patient Q data. From Eq. (3), it is evident that improvements in theoretical precision would require an improved SNR and/or an improved shimming to reduce the linewidths and, thus, damping. An increasing availability of higher field scanners and phased array coils is likely to help. A simple strategy with existing hardware would be to increase voxel dimensions. Even a modest increase from the 10 mm used in this study to 12 mm would result in an SNR improvement of 70% at the expense of slightly reduced
705
spatial resolution. However, field inhomogeneities would be poorer across larger voxels, thus increasing the linewidths. The stability of scanner electronics may still limit the achievable precision. In the stroke study, we found that DW lesion temperature overall was significantly higher than normal brain temperature. The small mean difference masks quite substantial differences in some patients, with extreme values ranging from N28C-higher temperatures to about 28C-lower temperatures between lesion and normal brain (Fig. 5). The reason for regional differences is unknown. They have not been documented previously due to lack of suitable techniques. Temperature elevation may be a consequence of ischemic metabolism, altered cerebral blood flow or inflammatory responses in ischemic tissue. Clearly, serial studies are required to fully investigate many of these issues. The 10-min scanning time for SI is acceptable within a standard clinical examination, but is excessive if repeated scans are required to monitor thermal intervention during examination. In that case, the phase difference method [11,12] might be more appropriate as it is faster and capable of higher spatial resolution. However, phase difference methods cannot yield absolute temperatures or temperature differences within the brain at one time point and, therefore, could not be used for the type of study reported here. Faster SI techniques [30] might allow multislice temperature measurements within the same scanning time. In summary, the temperature measurement technique described here was applied successfully in a study of ischemic stroke and could be used to study other brain disorders and to monitor therapies. Regional brain temperatures can be measured noninvasively and simultaneously with metabolite concentrations and other imaging sequences. Currently, however, the technique is limited to pooled data from group studies. More research is required to further evaluate the technique and to interpret the results. Future developments include the use of higher field scanners and phased array coils to improve SNR and, thus, the precision of temperature measurements.
Acknowledgments This study was funded by the UK Stroke Association (grant TSA 02/01) and was performed at the SHEFC Brain Imaging Research Centre for Scotland (www.dcn.ed.ac.uk/ bic). Dr. Karaszewski was supported by a bDr. James and Bozena Bain Memorial Trust Fund ScholarshipQ and a EUROGENDIS Marie Curie Scholarship. Dr. Wartolowska was supported by the Dlugolecka–Graham Trust, University of Edinburgh. Dr. Armitage was funded by the Row Fogo Charitable Research Foundation. Dr. Haga was funded by a Research into Aging Fellowship. We thank Drs. Steff Lewis and Francesca Chappell (Division of Clinical Neurosciences, University of Edinburgh) for statistical advice,
706
I. Marshall et al. / Magnetic Resonance Imaging 24 (2006) 699 – 706
and Prof. Martin Dennis for access to stroke patients. We thank Dr. Karine Macritchie (University of Newcastle) for access to healthy volunteers.
References [1] Olsen TS, Weber UJ, Kammersgaard LP. Therapeutic hypothermia for acute stroke. Lancet Neurol 2003;2:410 – 6. [2] Reith J, Jorgensen HS, Pedersen PM, Nakayama H, Raaschou HO, Jeppesen LL, et al. Body temperature in acute stroke: relation to stroke severity, infarct size, mortality, and outcome. Lancet 1996;347:422 – 5. [3] Castillo J, Davalos A, Marrugat J, Noya M. Timing for fever-related brain damage in acute ischemic stroke. Stroke 1998;29:2455 – 60. [4] Wass CT, Lanier WL, Hofer RE, Scheithauer BW, Andrews AG. Temperature changes of greater than or equal to 1 degree C alter functional neurologic outcome and histopathology in a canine model of complete cerebral ischemia. Anesthesiology 1995;83:325 – 35. [5] Zhang Y, Wong KC, Zhang Z. The effect of intraischemic mild hypothermia on focal cerebral ischemia/reperfusion injury. Acta Anaesthesiol Sin 2001;39:65 – 9. [6] Colbourne F, Corbett D. Delayed postischemic hypothermia: a six-month survival study using behavioral and histological assessments of neuroprotection. J Neurosci 1995;15:7250 – 60. [7] De Georgia MA. Cooling for Acute Ischemic Brain Damage (COOL AID): a feasibility trial of endovascular cooling. Neurology 2004;63:312 – 7. [8] Berger C, Schramm P, Schwab S. Reduction of diffusion-weighted MRI lesion volume after early moderate hypothermia in ischemic stroke. Stroke 2005;36:e56 – e58. [9] Schwab S, Spranger M, Aschoff A, Steiner T, Hacke W. Brain temperature monitoring and modulation in patients with severe MCA infarction. Neurology 1997;48:762 – 7. [10] Legos JJ. Programmable microchip monitoring of post-stroke pyrexia: effects of aspirin and paracetamol on temperature and infarct size in the rat. J Neurosci Methods 2002;113:159 – 66. [11] Germain D, Chevallier P, Laurent A, Saint-Jalmes H. MR monitoring of tumour thermal therapy. Magn Reson Materials Phys Biol Med (MAGMA) 2001;13:47 – 59. [12] Kuroda K, Chung AH, Hynynen K, Jolesz FA. Calibration of water proton chemical shift with temperature for non-invasive temperature imaging during focused ultrasound surgery. J Magn Reson Imaging 1998;8:175 – 81. [13] Cady EB, D’Souza PC, Penrice J, Lorek A. The estimation of local brain temperature by in vivo 1H magnetic resonance spectroscopy. Magn Reson Med 1995;33:862 – 7. [14] Corbett RJ, Laptook AR, Tollefsbol G, Kim B. Validation of a noninvasive method to measure brain temperature in vivo using 1H NMR spectroscopy. J Neurochem 1995;64:1224 – 30.
[15] Kuroda K, Mulkern RV, Oshio K, Panych LP, Nakai T, Moriya T, et al. Temperature mapping using the water proton chemical shift: selfreferenced method with echo-planar spectroscopic imaging. Magn Reson Med 2000;43:220 – 5. [16] Corbett RJ, Purdy PD, Laptook AR, Chaney C, Garcia D. Noninvasive measurement of brain temperature after stroke. AJNR Am J Neuroradiol 1999;20:1851 – 7. [17] Corbett R, Laptook A, Weatherall P. Noninvasive measurements of human brain temperature using volume-localized proton magnetic resonance spectroscopy. J Cereb Blood Flow Metab 1997;17:363 – 9. [18] Corbett RJ, Laptook AR. Failure of localized head cooling to reduce brain temperature in adult humans. Neuroreport 1998;9:2721 – 5. [19] Marshall I, Higinbotham J, Bruce S, Freise A. Use of Voigt lineshape for quantification of in vivo 1H spectra. Magn Reson Med 1997;37: 651 – 7. [20] Press WH, Vetterling WT, Teukolsky SA, Flannery BP. Numerical recipes in C. 2nd ed. Cambridge7 Cambridge University Press; 1992. p. 274 – 90. [21] van den Boogaart A, van Ormondt D, Pijnappel WWF, de Beer R, Ala-Korpela M. Removal of the water resonance from 1H magnetic resonance spectra. In: McWhirter JG, editor. Mathematics in signal processing III. Oxford7 Clarendon Press; 1994. p. 175 – 95. [22] Vanhamme L, van den Boogaart A, Van Huffel S. Improved method for accurate and efficient quantification of MRS data with use of priorknowledge. J Magn Reson 1997;129:35 – 43. [23] van den Boogaart A, Van Hecke A, Van Huffel P, GraveronDemilly S, van Ormondt D, de Beer R. MRUI: a graphical user interface for accurate routine MRS data analysis. Proceedings of the ESMRMB 13th Annual Meeting, Prague 1996. p. 318. [24] Klose U. In vivo proton spectroscopy in the presence of eddy currents. Magn Reson Med 1990;14:25 – 30. [25] Cramer H. Mathematical methods of statistics. Princeton7 University Press; 1946. [26] Rao CR. Minimum variance and the estimation of several parameters. Proc Camb Phil Soc 1946;43:280 – 3. [27] Cavassila S, Deval S, Huegen C, van Ormondt D, GraveronDemilly D. Cramer–Rao bound expressions for parametric estimation of overlapping peaks: influence of prior knowledge. J Magn Reson 2000;143:311 – 20. [28] Kuroda K, Suzuki Y, Ishihara Y, Okamoto K, Suzuki Y. Temperature mapping using water proton chemical shift obtained with 3D-MRSI: feasibility in vivo. Magn Reson Med 1996;35:20 – 9. [29] Cvoro V, Wartolowska K, Farrall AJ, Armitage PA, Marshall I, Bastin ME, et al. Voxel-based analysis of the relationship between metabolites, diffusion and perfusion parameters in acute ischemic stroke. Stroke 2004;35:E263. [30] Posse S, Dager SR, Richards TL, Yuan C, Ogg R, Artru AA, et al. In vivo measurement of regional brain metabolic response to hyperventilation using magnetic resonance: proton echo planar spectroscopic imaging (PEPSI). Magn Reson Med 1997;37:858 – 65.
Measurement of regional brain temperature using proton spectroscopic imaging: validation and application to acute ischemic stroke Ian Marshall4, Bartosz Karaszewski1, Joanna M. Wardlaw, Vera Cvoro, Karolina Wartolowska2, Paul A. Armitage, Trevor Carpenter, Mark E. Bastin, Andrew Farrall, Kristin Haga SHEFC Brain Imaging Research Centre for Scotland, University of Edinburgh, Edinburgh, UK Received 15 October 2005; accepted 5 February 2006
Abstract A magnetic resonance proton spectroscopic imaging (SI) technique was developed to measure regional brain temperatures in human subjects. The technique was validated in a homogeneous phantom and in four healthy volunteers. Simulations and calculations determined the theoretical measurement precision as approximately F0.38C for individual 1-ml voxels. In healthy volunteers, repeated measurements on individual voxels had an S.D. = 1.28C. In a clinical study, 40 patients with acute ischemic stroke were imaged within 26 h (mean, 10 h) of onset. Temperatures were highest in the region that appeared abnormal (i.e., ischemic) on diffusion-weighted imaging (DWI) compared with a normal-appearing brain. The mean temperature difference between the DWI blesionQ area and the b normal brain Q was 0.178C [ P b 103; range, 2.458C (hotter) – 2.178C (cooler)]. Noninvasive temperature measurement by SI has sufficient precision to be used in studies of pathophysiology in stroke and in other brain disorders and to monitor therapies. D 2006 Elsevier Inc. All rights reserved. Keywords: Spectroscopic imaging; Temperature; Brain; Stroke
1. Introduction Acute brain injury from a number of causes (e.g., stroke) is common and is associated with elevated body temperature [1–3], which in turn is associated with poor functional outcome. The cause of elevated body temperature and its relationship to brain temperature are unknown. A more severe stroke is associated with higher body temperatures [1,4]. In experimental models, cooling (hypothermia) may reduce lesion size [1,5] and improve outcome [1,6,7]. There is evidence that the same may be true clinically [8]. Direct
This work was presented, in part, at the 22nd Annual Meeting of ESMRMB, Basel, Switzerland, September 2005 (abstract 320). 4 Corresponding author. Medical Physics, Western General Hospital, EH4 2XU Edinburgh, UK. Tel.: +44 131 537 1661; fax: +44 131 537 1026. E-mail address: [email protected] (I. Marshall). 1 Current address: Department of Neurology of Adults, Medical University of Gdansk, Gdansk, Poland. 2 Current address: Department of Human Anatomy and Genetics, University of Oxford, Oxford, UK. 0730-725X/$ – see front matter D 2006 Elsevier Inc. All rights reserved. doi:10.1016/j.mri.2006.02.002
measurement of brain temperature might provide insights into the pathological processes taking place in ischemic and infarcted tissues [9]. Conventional methods of measuring brain temperature require invasive probes and can monitor only one or a few locations simultaneously [9,10]. Clearly, a reliable noninvasive method is highly desirable and could have many clinical uses. Several magnetic resonance parameters can be used for the noninvasive measurement of regional temperatures [11], including the diffusion coefficient, the longitudinal relaxation time constant (T 1) and the proton resonance frequency (PRF). Methods based on water PRF, which has a temperature dependence of approximately 0.01 ppm/8C [11], are the most popular. In bphase-shiftQ PRF techniques, the pixel-by-pixel phase change between successive gradient-echo images is used to monitor thermal interventions such as forced heating [11,12]. Although rapid and yielding spatial resolutions of imaging, these techniques are restricted to monitoring temperature changes and, thus, are not suitable for observational studies. The second class of PRF techniques uses spectroscopic acquisition to measure water
700
I. Marshall et al. / Magnetic Resonance Imaging 24 (2006) 699 – 706
chemical shift relative to a temperature-independent reference chemical shift [13 –15]. The reference compound depends on the application [e.g., lipids have been used for liver imaging [15], while N-acetyl (NA) compounds are most suitable for brain imaging [13,14]. These spectroscopic techniques are, in principle, capable of estimating absolute temperatures. Localized brain temperatures in experimental animals [14,16], healthy adults [17,18] and newborn infants [13] have been measured using single-voxel spectroscopy. However, the potential of spectroscopic imaging (SI) to map regional brain temperatures in diseases has not yet been exploited. Studies of brain temperature in stroke patients have so far been limited to single-point measurements using invasive sensors [9,10]. There have been no reports of brain temperature measurements made with magnetic resonance spectroscopic techniques in patients with acute ischemic stroke. Indeed, a theoretical problem for these techniques in ischemic stroke may be that the associated reduction in NA due to neuronal death may make it impossible to assess the relative water chemical shift. In this work, we validated a method of temperature measurement based on standard SI sequence in a test phantom and in four normal volunteers, and we used it in a study of stroke patients to measure temperature in an ischemic lesion [as defined on diffusion-weighted imaging (DWI)] and in a normal brain.
2. Methods 2.1. Theory of spectroscopic temperature measurement Temperature T is derived from the chemical shift of water (CSH2O) using a relation of the form: T ¼ Tref þ k ðCSH2 O CSref Þ
ð1Þ
where CSref is the (temperature-independent) chemical shift of a reference compound, k is the coefficient of proportionality and Tref is the reference temperature. In this work, scanner frequency was adjusted to bring the water signal onto resonance, and we constantly assigned a nominal value of 4.70 ppm to water frequency. The chemical shift of brain metabolites therefore appeared to vary with temperature, and a working version of Eq. (1) was given by: T ¼ 37 þ 100ðCSNA 2:035Þ
ð2Þ
where CSNA is the apparent chemical shift of NA compounds (expressed in ppm), the dominant singlet peak chosen for temperature determination in this work. A reference value of CSNA = 2.035 ppm was determined from occipital voxels in a group of 20 healthy control subjects (age, 46F10 years) with an assumed brain temperature of 378C. The constant of proportionality (1008C/ppm or, equivalently, 0.01 ppm/8C) was taken from the literature [11]. Eq. (2) was used in the simulation of
spectroscopic data and in the calculation of voxel temperatures from actual data. 2.2. Simulations Synthetic free induction decay (FID) signals were created to represent signals typically obtained during in vivo spectroscopic brain examinations on our 1.5-T scanner. FIDs consisted of 512 complex data points with a sample spacing of 1 ms and comprised a sum of exponentially decaying sinusoids corresponding to the major metabolites choline, creatine and NA compounds. Time domain amplitudes were 60, 60 and 100 U for the choline, creatine and NA components, respectively, and a common T 2 value of 300 ms was used. A bresidual water Q signal with an amplitude of 5000 U was added to represent the typical performance of our chemical-shift-selective (CHESS) water suppression across SI slices. FIDs were multiplied by a Gaussian envelope with a full width at half-maximum (FWHM) of 4 Hz to simulate line broadening due to intravoxel inhomogeneities [19]. In a first set of simulations (b temperature variation Q), the chemical shift of the metabolites (relative to water) was varied according to Eq. (2) to represent temperatures ranging from 368C to 398C in steps of 0.58C. At each chemical shift, a set of 100 FIDs was generated with Gaussian (i.e., normally distributed) noise [20] having an S.D. =50 U. This noise level is representative of 1-ml voxels on our scanner. In a second group of simulations (b NA amplitudeQ), the amplitude of the NA signal was varied from 100 to 20 U in steps of 10, with all other parameters being held constant. In a third set of simulations (bline broadening Q), the Gaussian line broadening was 4, 6, 8 or 10 Hz. Finally, the effect of noise level was investigated by generating a set of FIDs with a noise S.D. =20 U, with all other parameters being the same as for 378C FIDs in the first set of simulations. Groups of FID were processed in a standard manner. Briefly, the residual water component was first removed from each FID using the Hankel Lanczos Singular Value Decomposition technique [21]. Resulting signals were then quantified in the time domain using the AMARES algorithm [22] within the MRUI package [23]. A model consisting of three in-phase Gaussian components was used, as Gaussian lineshapes better describe in vivo data than Lorentzian lineshapes [19]. The mean and standard deviation of the NA peak chemical shift were determined for each group of FIDs and converted to equivalent temperatures using Eq. (2). 2.3. In vitro validation Initial in vitro experiments were carried out to determine measurement precision (standard deviation across the SI slice) under ideal conditions. A homogeneous phantom containing metabolites at physiological concentrations was imaged once a month as part of a quality assurance program. The phantom was kept in a scanner room to
I. Marshall et al. / Magnetic Resonance Imaging 24 (2006) 699 – 706
701
2.4. In vivo validation All human scannings were approved by the Local Research Ethics Committee. Initial in vivo evaluation was carried out on four healthy male volunteers (ages, 23– 38 years), who each had four SI scans within one examination. The SI slice was positioned at the level of the basal ganglia, and sequence parameters were the same as for the in vitro study. Volunteers were removed from the scanner and repositioned between the second and third scans. Spectroscopic data were processed as described above, and cerebral temperature for each valid voxel was calculated using Eq. (2). Statistical analysis was carried out using an autoregressive mixed model in SAS Version 9.1 (www.sas.com). 2.5. Stroke patients
Fig. 1. Determination of temperature from simulated FIDs representing in vivo signals from 1-ml voxels in the human brain at various temperatures. At each temperature, the mean and standard deviation are shown for a group of 100 FIDs with a Gaussian noise S.D. of 50 U, an NA level of 100 U and a Gaussian line broadening of 4 Hz. Actual simulated temperatures are shown by the line of identity (dashed line).
ensure thermal equilibrium, and its temperature was determined by a liquid crystal display (LCD) thermometer strip permanently attached to it. The reading accuracy was estimated as F18C. All magnetic resonance imaging (MRI) measurements were made on a 1.5-T Signa scanner (GE Healthcare, Slough, UK) fitted with standard head coil. The manufacturer’s single-slice SI sequence was used with PRESS excitation, an echo time (T E) of 145 ms and a repetition time (T R) of 1000 ms. The field of view (FOV) was 32 cm, with a phase-encoding grid of 2424 and a slice thickness of 10 mm. Automatic shimming and CHESS water suppression (with three pulses) were applied. For each phase encoding, 512 complex data points were acquired with a sampling interval of 1 ms. SI scanning time was approximately 10 min. Raw spectroscopic data were transferred to a workstation and interpolated to a 3232 matrix during Fourier transformation, thereby yielding voxels of nominal dimensions of 101010 mm. Zero-order phase correction was carried out on a voxel-by-voxel basis using residual water signals [24], thereby effectively bringing water to a fixed nominal chemical shift of 4.70 ppm. Subsequent processing and AMARES quantification were as described above. Spectra were automatically discarded if fitted linewidths were b1 Hz or N 10 Hz, or if metabolite peaks were N0.1 ppm offset of their expected values, or if voxels lay on the edges of the PRESS excitation region. All spectra were also inspected visually and discarded if judged to be of poor quality (e.g., having a badly elevated baseline or containing spurious peaks). Temperatures for each remaining voxel were calculated using Eq. (2).
We recruited 40 patients (mean age, 78 years; range, 58– 95 years) with symptoms of moderate to severe cortical ischemic stroke in an MRI study to compare metabolites in terms of their appearance on DWI. The study was approved by the Local Research Ethics Committee, and all subjects gave written informed consent. Patients had magnetic resonance brain examination within 26 h (mean, 10 h) of stroke onset. The examination included axial diffusionweighted spin-echo echo-planar imaging (DWI) and SI centered on the slice showing the maximum ischemic lesion extent on DWI. The imaging parameters for DWI were: FOV=240 mm, axial slice= 15, thickness = 5 mm, slice gap =1 mm, acquisition matrix = 128128 and T R/ T E = 10 s/97 ms. Diffusion-sensitizing gradients with scalar b values of 1000 s/mm2 were applied in six noncollinear directions. The total examination time, including patient setup, was typically 40 –50 min. SI parameters and data
Fig. 2. Determination of temperature from simulated FIDs representing in vivo signals from 1-ml voxels in the human brain with various levels of NA compounds. At each level, the mean and standard deviation are shown for a group of 100 FIDs with Gaussian noise S.D. of 50 U and a Gaussian line broadening of 4 Hz. The actual simulated temperature is 37.08C (dashed line).
702
I. Marshall et al. / Magnetic Resonance Imaging 24 (2006) 699 – 706
Fig. 3. (A) Mean and (B) standard deviation temperature maps of four successive measurements made at the level of the basal ganglia in a healthy volunteer. The temperature scale is given in degrees Celsius.
processing were as described above, except that a fivecomponent model was used for AMARES quantification. Prior knowledge was applied to link the two additional Table 1 Effect of Gaussian line broadening on temperatures determined from simulated FIDs representing in vivo signals from 1-ml voxels in the human brain Applied line broadening (Hz)
Temperature (meanFS.D.)
Fitted FWHM (Hz)
4 6 8 10
37.02F0.29 37.01F0.48 36.99F0.67 36.97F0.87
4.7 6.7 8.6 10.5
For each line broadening, the mean and standard deviation are shown for a group of 100 FIDs with a Brownian noise S.D. = 50 U, NA amplitude = 100 U, intrinsic T 2 = 300 ms and simulated temperature = 37.08C. Also shown is the Gaussian linewidth (FWHM) fitted by AMARES quantification.
Fig. 4. (A) Diffusion-weighted image and (B) corresponding temperature map of a patient with an acute ischemic stroke lesion in the left parietal region. In (B), the lesion outline is shown in red. The temperature scale is given in degrees Celsius.
components in order to model the doublet signal of lactate. Cerebral temperature for each valid voxel was calculated using Eq. (2). The SI voxel grid was superimposed on the DWI image. SI voxels overlying the DWI lesion were identified as b LES.Q Voxels containing a normal-appearing brain located beyond the DWI lesion were identified as b ipsilateral normal Q (INL) or b contralateral normal Q (CNL) in the ipsilateral and contralateral hemispheres, respectively. Voxels covering the cerebrospinal fluid were excluded. For each patient, mean temperatures were calculated for each tissue category, and mean temperature differences between categories were calculated. Finally, the group means of these values were calculated across all patients. Statistical analysis was carried out using odds ratios by fixed-effects
I. Marshall et al. / Magnetic Resonance Imaging 24 (2006) 699 – 706
703
model in RevMan software (http://www.cc-ims.net/ RevMan). This analysis was primarily within patients [i.e., comparing abnormal vs. normal temperature within each patient, thereby controlling for external factors (such as variation in scanner performance) that might vary between patients].
3. Results 3.1. Simulations In the first set of simulations, temperatures deduced from the quantification of synthetic FIDs were in excellent agreement with actual temperatures (Fig. 1). At each temperature, there was no bias in mean results, and the standard deviation of results was 0.38C. In the second group of simulations, the mean temperature was accurately determined at each level of NA (Fig. 2). The standard deviation increased in inverse proportion to the level of NA, as NA fell from 100 to 40 U. Below 40 U, the standard deviation increased more rapidly than in the inverse proportion. At an NA level of 20 U, the fitting algorithm failed in 5 of 100 cases, and the standard deviation of the remaining 95 cases was 2.58C, compared with a value of 1.58C extrapolated from higher levels and assuming an inverse linear relationship. The quality of these spectra was such that they would have been rejected in an experimental situation. The effect of Gaussian line broadening is shown in Table 1. The temperature was accurately determined for all line broadenings studied, while the standard deviation increased with increasing broadening. Reducing the noise standard deviation from 50 to 20 U reduced the standard deviation of temperature estimations from 0.308C to 0.128C. 3.2. In vitro validation Each set of voxel temperatures estimated from monthly SI measurements of the phantom had an S.D. b 0.28C. Over the observed seasonal range of 19–248C, the difference (meanFS.D.) between derived temperatures and LCD thermometer temperatures was 0.4F0.78C, which is within the reading accuracy of the LCD thermometer.
Table 2 Summary of regional brain temperature differences in a study of 40 patients with acute ischemic stroke Groups
Temperature difference (95% confidence interval) (8C)
n
LES versus INL+CNL LES versus INL LES versus CNL CNL versus INL
0.17 0.50 0.38 0.22
40 24 29 35
(0.07, (0.34, (0.27, (0.12,
0.27) 0.66) 0.49) 0.32)
All findings are statistically significant ( P b 103). n, number of contributing patients (see also Fig. 5).
Fig. 5. Histograms of intraindividual temperature differences (8C) between LES, INL and CNL brain tissues in a study of 40 patients with acute ischemic stroke (see also Table 2).
3.3. In vivo validation Useful data were obtained from approximately 30 voxels (mean = 29, S.D. =6) in each cerebral hemisphere of each volunteer on each scan. Spectral quality was low in the frontal region. The overall mean brain temperature (all subjects, all scans) was 36.58C. There was no significant difference between hemispheres. There was a small but significant ( P b.001) effect of scan number, with the temperature falling by 0.098C between successive scans. However, this effect was not significant for Subject 1 alone; therefore, we combined separate scan data for this subject for the purposes of Fig. 3. The statistical model gave a residual variance of 1.48C (i.e., repeated measurements on individual voxels had an S.D. = 1.28C). 3.4. Stroke patients Most spectra were of sufficient quality, and most lesions were large enough for each patient to yield temperature readings from at least two voxels (and, in some cases, more than 40 voxels) in each of the tissue groups (LES, INL and CNL). A typical example is shown in Fig. 4. Although ischemic stroke reduces NA levels, spectral peak areas were still sufficient to allow valid temperature measurements from many voxels in all patients. The group mean temperatures of different tissue classes were not significantly
704
I. Marshall et al. / Magnetic Resonance Imaging 24 (2006) 699 – 706
different. However, the group means of intraindividual temperature differences were highly significant ( P b103) (Table 2). Histograms of the differences are shown in Fig. 5. Lesion temperature was 0.178C higher than normal tissue (INL+CNL combined) temperature [range, 2.458C (hotter)–2.178C (cooler)], and CNL tissue temperature was 0.228C higher than INL tissue temperature [range, 1.088C (hotter)–1.568C (cooler)]. 4. Discussion Temperatures were estimated accurately for all simulated FIDs over the range of temperatures, NA levels and linewidths expected in vivo. Under ideal conditions (homogeneous spectroscopic phantom at constant uniform temperature), the SI technique yielded voxel temperatures with an S.D. = 0.28C across the measurement slice. Temperature determination using proton spectroscopic techniques requires estimation of the frequency of water and reference peaks. Given that the scanner will adjust to water resonance frequency, the problem reduces to that of determining the frequency of the reference peak. We used the singlet peak of NA for this work. Simulations suggested a frequency precision of 0.003 ppm and, thus, a temperature precision of 0.38C for typical 1-ml voxels in healthy volunteers. These simulations assumed a 4-Hz Gaussian line broadening of an NA signal having an intrinsic T 2 =300 ms and an amplitude =100 U, corrupted with noise S.D. = 50 U. A framework for the estimation of model parameters is provided by the Cramer–Rao (CR) theory [25,26]. The theory is based on the sensitivity of a model function to constituent parameters and can be used to evaluate the achievable precision of parameter estimates given the constraints of data and assuming bperfect Q quantification. The resulting CR boundary (CRB) is the best that can be achieved. An analytical expression of CRB on the frequency determination for a single Lorentzian peak has been given [27] as: pffiffiffi pffiffiffi r CRBN ¼ 2 2ð aÞ3=2 ts ð3Þ c where a is damping, t s is sampling interval, r is noise standard deviation and c is signal amplitude. All quantities relate to the time domain. Note that the ratio in parentheses is the inverse of the signal-to-noise ratio (SNR). For a Lorentzian lineshape, damping is related to FWHM by a =p FWHM. Evaluating Eq. (3) using the NA and noise parameters given above, together with a sampling interval of 1 ms and a FWHM of 4 Hz, gives CRBN =2.0 rad or 0.3 Hz. At a field strength of 1.5 T, this is equivalent to 0.005 ppm or 0.58C. This result is in reasonable agreement with the value of 0.38C found from simulations. The discrepancy is due to the fact that Eq. (3) holds strictly only for Lorentzian lineshapes (i.e., exponentially decaying FIDs) and will not be accurate for Gaussian and mixed lineshapes used in simulations and found in in vivo data
[19]. Corresponding expressions for lineshapes other than Lorentzian are not available, but Lorentzian calculation is offered here as a guide. Intuitively, we might expect a b worseQ result for Lorentzians than Gaussians, as a significant part of the energy of Lorentzian lineshapes is contained in extensive tails that are difficult to detect above the noise level. Eq. (3) shows that CRBN increases more than linearly with linewidth, and, thus, shimming is important in achieving precise temperature estimation. Temperature standard deviations in Table 1 increased with linewidth to a power of approximately 1.3. A temperature resolution of 0.38C requires a 0.003-ppm frequency resolution, which is 0.2 Hz at 1.5 T. To measure such a small difference in a conventional manner from a spectrum (i.e., a Fourier-transformed FID) would require an FID duration of 5 s, or 2.5 s if zero filling were applied. Given that intrinsic T 2 decay and T 2* processes limit the usable signal duration to a few hundred milliseconds at best, this is clearly not practical. A more sophisticated method is required for frequency determination from a limited duration signal. In this work, we used AMARES for the time domain modeling of complex FIDs. Effectively, the change in phase between successive data points is used to estimate the corresponding frequency. Modeling in the frequency domain is also possible [28], and both methods are, in principle, capable of giving results limited only by CR bounds. Temperature determination using proton spectroscopic techniques requires that both water peak and internal reference peak be visible in each spectrum. Spectroscopic data are normally acquired with water suppression so as not to swamp weak metabolite signals with overwhelming signal due to water. In clinical SI, there is insufficient time to also collect a nonsuppressed data set and, therefore, residual water signal must be large enough for frequency (hence, temperature) determination. In the standard implementation of SI that we used, we found that even with water suppression, the residual water signal was typically 30–50 times greater than metabolite signals. We did not need to deliberately b degrade Q water suppression. Of course, the selective RF pulse(s) used for water suppression should not distort the shape of the (residual) water peak; otherwise, an artifactual b shift, Q which is misinterpreted as a temperature difference, may be created. In the worst case, this effect could vary across the FOV if field inhomogeneities shifted the primary water resonance toward either edge of the frequency band of the pulse(s). In phantom experiments comparing suppressed and nonsuppressed water acquisition (data not shown), we were unable to detect any systematic differences in estimated temperatures. We are therefore satisfied that our standard water suppression technique is valid. We used the residual water signal for eddy current and zero-order phase correction [24]. We used the NA peak as the temperature-independent reference. NA is the dominant singlet peak in proton brain spectra acquired over a wide range of echo times and is, therefore,
I. Marshall et al. / Magnetic Resonance Imaging 24 (2006) 699 – 706
the obvious candidate. Even within DWI-infarcted (LES) tissue, we found that NA levels were typically 66% [29] and usually no lower than 30% of the level found in healthy brain tissue. This finding could be explained by partial volume effect, or by the fact that we imaged patients sufficiently soon after stroke while metabolite changes were still evolving. Whatever the explanation, we did not need to discard voxels solely because of insufficient NA signal. The slice selection gradient strength in our SI sequence was 6 mT/m for the 10-mm slice thickness used. With a nominal chemical shift difference of 2.7 ppm between water and NA, this leads to a slice position offset of 0.7 mm. We consider this theoretical offset to be negligible, as it is well within the uncertainty in slice prescription and subject head motion during a typical examination. The simulations and CR calculations suggest a theoretical temperature precision of 0.38C for individual wellshimmed 1-ml voxels in healthy volunteers. The residual S.D. of 1.28C for repeated measurements on individual voxels in four healthy volunteers was higher than what might be suggested by theoretical precision. Increased random variation may be due to subject motion, spectral contamination from lipids or scanner instabilities, and sets a realistic bound on what can presently be achieved in vivo. We hypothesize that the fall in temperature by 0.098C between successive scans was due to the combined effects of subject inactivity and cooling by the air flow in the scanner bore. Measurement repeatability is likely to be worse in patients when voxels with NA levels down to approximately 30% of normal levels might be accepted as valid. Thus, intraindividual temperature differences from several subjects need to be pooled in order to detect differences on the order of 0.18C, as reported here. It is too early to say whether the method can measure temperature in individuals. However, it is possible that it could, as temperature measurements are available in absolute units. To overcome concerns regarding precision and repeatability and the problem of different numbers of voxels contributing to the analysis in each patient, we have concentrated our analysis on within-patient differences so that each subject’s normal tissue acts as a control. This approach minimizes the effects of between-session variability (as yet not fully quantified for this technique) and has allowed us to obtain highly significant P values despite small differences in temperature. To further err on the side of caution, we have used grouped bwithin-patient Q data. From Eq. (3), it is evident that improvements in theoretical precision would require an improved SNR and/or an improved shimming to reduce the linewidths and, thus, damping. An increasing availability of higher field scanners and phased array coils is likely to help. A simple strategy with existing hardware would be to increase voxel dimensions. Even a modest increase from the 10 mm used in this study to 12 mm would result in an SNR improvement of 70% at the expense of slightly reduced
705
spatial resolution. However, field inhomogeneities would be poorer across larger voxels, thus increasing the linewidths. The stability of scanner electronics may still limit the achievable precision. In the stroke study, we found that DW lesion temperature overall was significantly higher than normal brain temperature. The small mean difference masks quite substantial differences in some patients, with extreme values ranging from N28C-higher temperatures to about 28C-lower temperatures between lesion and normal brain (Fig. 5). The reason for regional differences is unknown. They have not been documented previously due to lack of suitable techniques. Temperature elevation may be a consequence of ischemic metabolism, altered cerebral blood flow or inflammatory responses in ischemic tissue. Clearly, serial studies are required to fully investigate many of these issues. The 10-min scanning time for SI is acceptable within a standard clinical examination, but is excessive if repeated scans are required to monitor thermal intervention during examination. In that case, the phase difference method [11,12] might be more appropriate as it is faster and capable of higher spatial resolution. However, phase difference methods cannot yield absolute temperatures or temperature differences within the brain at one time point and, therefore, could not be used for the type of study reported here. Faster SI techniques [30] might allow multislice temperature measurements within the same scanning time. In summary, the temperature measurement technique described here was applied successfully in a study of ischemic stroke and could be used to study other brain disorders and to monitor therapies. Regional brain temperatures can be measured noninvasively and simultaneously with metabolite concentrations and other imaging sequences. Currently, however, the technique is limited to pooled data from group studies. More research is required to further evaluate the technique and to interpret the results. Future developments include the use of higher field scanners and phased array coils to improve SNR and, thus, the precision of temperature measurements.
Acknowledgments This study was funded by the UK Stroke Association (grant TSA 02/01) and was performed at the SHEFC Brain Imaging Research Centre for Scotland (www.dcn.ed.ac.uk/ bic). Dr. Karaszewski was supported by a bDr. James and Bozena Bain Memorial Trust Fund ScholarshipQ and a EUROGENDIS Marie Curie Scholarship. Dr. Wartolowska was supported by the Dlugolecka–Graham Trust, University of Edinburgh. Dr. Armitage was funded by the Row Fogo Charitable Research Foundation. Dr. Haga was funded by a Research into Aging Fellowship. We thank Drs. Steff Lewis and Francesca Chappell (Division of Clinical Neurosciences, University of Edinburgh) for statistical advice,
706
I. Marshall et al. / Magnetic Resonance Imaging 24 (2006) 699 – 706
and Prof. Martin Dennis for access to stroke patients. We thank Dr. Karine Macritchie (University of Newcastle) for access to healthy volunteers.
References [1] Olsen TS, Weber UJ, Kammersgaard LP. Therapeutic hypothermia for acute stroke. Lancet Neurol 2003;2:410 – 6. [2] Reith J, Jorgensen HS, Pedersen PM, Nakayama H, Raaschou HO, Jeppesen LL, et al. Body temperature in acute stroke: relation to stroke severity, infarct size, mortality, and outcome. Lancet 1996;347:422 – 5. [3] Castillo J, Davalos A, Marrugat J, Noya M. Timing for fever-related brain damage in acute ischemic stroke. Stroke 1998;29:2455 – 60. [4] Wass CT, Lanier WL, Hofer RE, Scheithauer BW, Andrews AG. Temperature changes of greater than or equal to 1 degree C alter functional neurologic outcome and histopathology in a canine model of complete cerebral ischemia. Anesthesiology 1995;83:325 – 35. [5] Zhang Y, Wong KC, Zhang Z. The effect of intraischemic mild hypothermia on focal cerebral ischemia/reperfusion injury. Acta Anaesthesiol Sin 2001;39:65 – 9. [6] Colbourne F, Corbett D. Delayed postischemic hypothermia: a six-month survival study using behavioral and histological assessments of neuroprotection. J Neurosci 1995;15:7250 – 60. [7] De Georgia MA. Cooling for Acute Ischemic Brain Damage (COOL AID): a feasibility trial of endovascular cooling. Neurology 2004;63:312 – 7. [8] Berger C, Schramm P, Schwab S. Reduction of diffusion-weighted MRI lesion volume after early moderate hypothermia in ischemic stroke. Stroke 2005;36:e56 – e58. [9] Schwab S, Spranger M, Aschoff A, Steiner T, Hacke W. Brain temperature monitoring and modulation in patients with severe MCA infarction. Neurology 1997;48:762 – 7. [10] Legos JJ. Programmable microchip monitoring of post-stroke pyrexia: effects of aspirin and paracetamol on temperature and infarct size in the rat. J Neurosci Methods 2002;113:159 – 66. [11] Germain D, Chevallier P, Laurent A, Saint-Jalmes H. MR monitoring of tumour thermal therapy. Magn Reson Materials Phys Biol Med (MAGMA) 2001;13:47 – 59. [12] Kuroda K, Chung AH, Hynynen K, Jolesz FA. Calibration of water proton chemical shift with temperature for non-invasive temperature imaging during focused ultrasound surgery. J Magn Reson Imaging 1998;8:175 – 81. [13] Cady EB, D’Souza PC, Penrice J, Lorek A. The estimation of local brain temperature by in vivo 1H magnetic resonance spectroscopy. Magn Reson Med 1995;33:862 – 7. [14] Corbett RJ, Laptook AR, Tollefsbol G, Kim B. Validation of a noninvasive method to measure brain temperature in vivo using 1H NMR spectroscopy. J Neurochem 1995;64:1224 – 30.
[15] Kuroda K, Mulkern RV, Oshio K, Panych LP, Nakai T, Moriya T, et al. Temperature mapping using the water proton chemical shift: selfreferenced method with echo-planar spectroscopic imaging. Magn Reson Med 2000;43:220 – 5. [16] Corbett RJ, Purdy PD, Laptook AR, Chaney C, Garcia D. Noninvasive measurement of brain temperature after stroke. AJNR Am J Neuroradiol 1999;20:1851 – 7. [17] Corbett R, Laptook A, Weatherall P. Noninvasive measurements of human brain temperature using volume-localized proton magnetic resonance spectroscopy. J Cereb Blood Flow Metab 1997;17:363 – 9. [18] Corbett RJ, Laptook AR. Failure of localized head cooling to reduce brain temperature in adult humans. Neuroreport 1998;9:2721 – 5. [19] Marshall I, Higinbotham J, Bruce S, Freise A. Use of Voigt lineshape for quantification of in vivo 1H spectra. Magn Reson Med 1997;37: 651 – 7. [20] Press WH, Vetterling WT, Teukolsky SA, Flannery BP. Numerical recipes in C. 2nd ed. Cambridge7 Cambridge University Press; 1992. p. 274 – 90. [21] van den Boogaart A, van Ormondt D, Pijnappel WWF, de Beer R, Ala-Korpela M. Removal of the water resonance from 1H magnetic resonance spectra. In: McWhirter JG, editor. Mathematics in signal processing III. Oxford7 Clarendon Press; 1994. p. 175 – 95. [22] Vanhamme L, van den Boogaart A, Van Huffel S. Improved method for accurate and efficient quantification of MRS data with use of priorknowledge. J Magn Reson 1997;129:35 – 43. [23] van den Boogaart A, Van Hecke A, Van Huffel P, GraveronDemilly S, van Ormondt D, de Beer R. MRUI: a graphical user interface for accurate routine MRS data analysis. Proceedings of the ESMRMB 13th Annual Meeting, Prague 1996. p. 318. [24] Klose U. In vivo proton spectroscopy in the presence of eddy currents. Magn Reson Med 1990;14:25 – 30. [25] Cramer H. Mathematical methods of statistics. Princeton7 University Press; 1946. [26] Rao CR. Minimum variance and the estimation of several parameters. Proc Camb Phil Soc 1946;43:280 – 3. [27] Cavassila S, Deval S, Huegen C, van Ormondt D, GraveronDemilly D. Cramer–Rao bound expressions for parametric estimation of overlapping peaks: influence of prior knowledge. J Magn Reson 2000;143:311 – 20. [28] Kuroda K, Suzuki Y, Ishihara Y, Okamoto K, Suzuki Y. Temperature mapping using water proton chemical shift obtained with 3D-MRSI: feasibility in vivo. Magn Reson Med 1996;35:20 – 9. [29] Cvoro V, Wartolowska K, Farrall AJ, Armitage PA, Marshall I, Bastin ME, et al. Voxel-based analysis of the relationship between metabolites, diffusion and perfusion parameters in acute ischemic stroke. Stroke 2004;35:E263. [30] Posse S, Dager SR, Richards TL, Yuan C, Ogg R, Artru AA, et al. In vivo measurement of regional brain metabolic response to hyperventilation using magnetic resonance: proton echo planar spectroscopic imaging (PEPSI). Magn Reson Med 1997;37:858 – 65.