Determination of regional brain temperature using proton magnetic resonance spectroscopy to assess brain–body temperature differences in healthy human subjects



Proton magnetic resonance spectroscopy (1H MRS) was used to determine brain temperature in healthy volunteers. Partially water-suppressed 1H MRS data sets were acquired at 3T from four different gray matter (GM)/white matter (WM) volumes. Brain temperatures were determined from the chemical-shift difference between the CH3 of N-acetyl aspartate (NAA) at 2.01 ppm and water. Brain temperatures in 1H MRS voxels of 2 × 2 × 2 cm3 showed no substantial heterogeneity. The volume-averaged temperature from single-voxel spectroscopy was compared with body temperatures obtained from the oral cavity, tympanum, and temporal artery regions. The mean brain parenchyma temperature was 0.5°C cooler than readings obtained from three extra-brain sites (P < 0.01). 1H MRS imaging (MRSI) data were acquired from a slice encompassing the single-voxel volumes to assess the ability of spectroscopic imaging to determine regional brain temperature within the imaging slice. Brain temperature away from the center of the brain determined by MRSI differed from that obtained by single-voxel MRS in the same brain region, possibly due to a poor line width (LW) in MRSI. The data are discussed in the light of proposed brain–body temperature gradients and the use of 1H MRSI to monitor brain temperature in pathologies, such as brain trauma. Magn Reson Med 57:59–66, 2007. © 2006 Wiley-Liss, Inc.

Our current knowledge about brain temperature in healthy humans is limited. It is generally assumed, however, that the brain is “hotter” than the rest of the body because of its high rate of cerebral metabolic activity and blood flow (1). This is borne out by theoretical models that suggest that the temperature of the brain is slightly higher than that of the body (2). Once the brain is damaged, whether by stroke or physical trauma (head injury), a dissociation commonly occurs between brain and body temperatures. The weight of the evidence indicates that average brain temperatures rise above body temperatures (3), typically by an average of 1°C, but by as much as 2°C in some cases (4, 5). Although the significance of a large, positive brain–body temperature gradient is currently unclear, there is some evidence that if the opposite occurs (i.e., brain temperature falls below body temperature) the prognosis is grave (4, 6–8). Thus, despite the prevailing view that damaged brain has a higher temperature than other body sites, this is clearly not the case for all patients at all times (6).

In clinical practice, brain temperature measurement is, of necessity, restricted to a single focal area. In recent years the traditional methods of intracerebral monitoring via the ventricles (9) has largely been replaced by the placement of intraparenchymal sensors directly into white matter (WM). However, the assumption that focal temperature reflects the temperature in other brain regions has yet to be confirmed.

Previous studies have measured local brain temperature using noninvasive techniques (10, 11). NMR-based techniques in both imaging and spectroscopy modes play a central role in noninvasive temperature mapping. The use of the proton (1H) chemical shift of water, referenced to a variety of endogenous metabolites to probe absolute brain temperature, overcomes the limitations of focal monitoring; however, this technique may lack the resolution required to identify small differences in temperature (12). N-acetylaspartate (NAA), an abundant brain chemical that is detectable in 1H brain spectra, provides a more precise reference compared to other endogenous substances, such as choline-containing metabolites and total creatine, due to a large NMR peak from NAA in healthy adult brain (10, 12). The chemical shift of the NAA resonance at 2.01 ppm within the physiological temperature range expected during health and critical illness is independent of pH and is physiologically stable. The water chemical shift has an almost linear dependence on temperature (10, 11). The 1H chemical-shift difference of the resonances between water and CH3 of NAA at 2.01 ppm can be modeled as an analytical mathematical function, and the differential provides adequate temperature resolution for noninvasive quantification of human brain temperature.

To better understand the significance of brain–body temperature differences reported after severe brain damage in humans, in the present study we investigated regional variations in local brain temperature in healthy subjects under normal physiological conditions, and assessed differences between local brain and body temperatures.


Subject Protocol

Following approval from the Birmingham University Research Ethics Committee, eight healthy subjects (three females and five males, 23–52 years old, median = 28 years) with no known history of brain injury or disease were recruited to the study. The subjects were studied in light outdoor clothing and were instructed to refrain from eating or drinking for 60 min before attending the MRI suite. Ten minutes before the subjects entered the magnet, their body temperature was measured at two conventional sites: the oral cavity (Toral) and the tympanum (Ttymp). A third body temperature measurement was obtained using an infrared scanner (Exergen; MA, USA) applied to the skin overlying the course of the temporal artery (Tt.a). The subjects lay supine and the head was supported by padding and foam and positioned centrally within the coil to ensure minimal rotation in any plane. The padding used to position and secure the head for MR scanning did not raise the local skin temperature during the time used for examination, as confirmed experimentally (not shown). On completion of the scan, repeated oral, tympanic, and temporal artery temperature measurements were made.

MRI and Spectroscopy

MRI and MR spectroscopy (MRS) measurements were performed at 3T using a Phillips Achieva scanner (Phillips Medical Inc., Best, The Netherlands) with a standard transmit-receive quadrature head coil. Axial T2-weighted images (field of view (FOV) = 230 mm, time to repetition (TR) = 3260 ms, time to echo (TE) = 80 ms, matrix size = 265 × 196, slice thickness = 4 mm, slice gap = 5 mm, 21 slices) were acquired for anatomical reference and for positioning of 1H MRS single-voxel volumes and the MRS imaging (MRSI) slice. Single-voxel 1H MR spectra were acquired (number of averages 32 preceded by four dummy prescans, total acquisition time = 54 s) using a double spin-echo method with acquisition parameters as follows: TR = 1500 ms, TE = 60 ms, sweep width (SW) =1500 Hz, 2 × 2 × 2 cm3 volume, and 1 k data points. Water suppression power at the water signal frequency was set to either 50% (all four anatomical locations) or 0% (middle of the brain, voxel 4; Fig. 1) of the power needed to remove the water magnetization. Spectra were recorded from four representative brain regions within gray matter (GM)/WM and deep WM of the surface and deep frontal lobe, respectively, and WM of the temporal and occipital lobes close to the middle of the brain (Fig. 1).

Figure 1.

Axial T2-weighted image at the level of the brain just above the lateral ventricles. Figure shows the four regions (boxes) in which single-voxel 1H MRS spectra were acquired. All spectra were acquired with 50% water suppression. The white grid represents the MRSI voxels. The imaging parameters are given in the text.

In addition, we acquired 1H MRSI data sets (total acquisition time = 3 min 20 s) from one slice (20 mm thick) positioned slightly above the level of the lateral ventricles to closely match the areas from which single-voxel spectra were acquired. The slice was shimmed for B0 homogeneity using both linear and second-order shims with at least two iterations. The outer volume saturation (OVS) was optimized and the level of 50% water saturation pulse power was determined. MRSI was performed using a double-echo method with FOV = 230 mm, TR = 1500 ms, TE = 60 ms, SW = 2000 Hz, and 1 k data points, to produce a 16 × 16 MRSI data matrix (nominal voxel dimensions 1.44 × 1.44 × 2 cm3 = 4.1 cm3) with eight scans for each phase-encoding step. A solution containing 12 mM NAA in phosphate-buffered saline (PBS, pH 7.1) in a square container (150 × 114 × 35 mm3) was scanned with both single-voxel 1H MRS and MRSI using the acquisition parameters as for human volunteers.

Both the MRS and MRSI data were zero-filled to 2 k data points before Fourier transformation (FT) without apodization filters. The MRS images were reconstructed using Philips software. The single-voxel and MRSI spectra were transferred from the scanner to a personal computer in spectral form. The single-voxel spectra were then analyzed by the line-shape fitting method, whereas the MRSI spectra were first transformed into free induction decays (FIDs) before an inverse fast FT (FFT) was conducted. Before each new FFT was performed, the FIDs were processed by an exponential function corresponding to 2 Hz line-broadening.

The volumes from which the brain temperature maps were generated were well within the brain tissue and excluded structures close to the surface of the brain. MRSI spectra in the superficial brain were often contaminated by residual lipid peaks, which hampered determination of the NAA resonance frequency and/or distorted the baseline. Therefore, these volumes were excluded from further analysis.

Line-Shape Fitting and Accuracy

In MRS it is common to assume a Lorentzian line shape based on a one-exponential transverse relaxation behavior of the spins. Here the real part, L, of the signals was used to estimate the spectral parameters in a line-shape fitting analysis. This part of the model function can be written in the frequency domain:

equation image(1)

where ai is the line width (LW) at half height, Ii is the intensity, νi is the resonance frequency, and ϕi is the phase angle of Lorentzian. The real parts of the water and NAA signals were fitted separately by using this line-shape model. The spectral regions covered 0.911 ppm (116.6 Hz) for water and 0.418 ppm (53.5 Hz) for NAA. The baseline variation was taken into account by a linear correction in the fitting analysis. This procedure yielded a good approximation for the baseline, especially in the cases of the single-voxel spectra (Fig. 2). The spectral parameters ai, Ii,, νI, and ϕi and the linear baselines co + c1ν for both MRS signals were automatically adjusted in the analysis of their spectral characteristics.

Figure 2.

Typical resonances in the four brain regions of Fig. 1: (a) voxel 1, (b) voxel 2, (c) voxel 3, and (d) voxel 4. exp, experimental spectra; calc, calculated spectra; diff, experimental minus calculated spectra showing the residuals. Acquisition parameters: TR = 1500 ms, TE = 60 ms, SW = 1500 Hz, 2 × 2 × 2 cm3 volume, 50% water suppression.

The observed 1H chemical-shift difference ΔH2O-NAA between signals of water and NAA was converted into temperature units using calibration data from Corbett et al. (11):

equation image(2)

The pre-processing, line-shape fitting analysis, and temperature conversion were performed offline with in-house-developed automated software written under a Matlab platform (Mathworks, Natick, MA, USA).

The errors of the temperatures for both single-voxel 1H MRS and MRSI methods were estimated using a homogenous phantom containing NAA (see above for phantom composition). Ten single-voxel MRS spectra from different locations of the phantom and 20 voxels from a 1H MRSI data set were used to determine temperature variation.

Measurement of Signal-to-Noise Ratio (SNR) and Simulation of Spectral Data Sets

Spectral quality influences the accuracy of resonance frequency determination and thus the precision of temperature quantification by 1H MRS. The key spectral characteristics affecting the resonance frequency include the SNR, LW, and digital resolution. We determined the SNR of the in vivo spectra for the NAA peak by dividing the peak height by the standard deviation (SD) of noise at the chemical-shift region of 0 ppm. LWs at half peak height for water and NAA peaks were determined from single-voxel and MRSI spectra.

To estimate the contributions of individual spectral characteristics to the precision of temperature determination, we simulated 1H MRS spectra to closely match those recorded in vivo using Lorentzian line shapes at different frequencies, intensities, and LWs for the water and NAA signals. The Lorentzian line shape was also used for the water line even though the resonance has non-Lorentzian features in vivo. However, as shown experimentally (see below), the water lines obtained in vivo are symmetrical and therefore the frequency can be accurately determined with the selected spectral analysis procedure (see above). Baseline and noise were taken from the experimental single-voxel data set. Simulated spectra with a precisely known frequency difference between water and NAA were generated to assess their individual effects on the accuracy of temperature determination by 1H MRS as follows: LW was varied from 1 to 15 Hz using a fixed SNR of 20 for NAA and a digital resolution of 1.5 Hz/point (typical values in single-voxel 1H MRS); SNR for NAA varied between 10 and 15 using fixed values for LW of 6 Hz and digital resolution of 1.5 Hz/point, and digital resolution varied from 0.75 to 2.75 Hz/point at fixed SNR of 20 for NAA and LW of 6 Hz. The simulated spectra for each variable were analyzed by the line-shape fitting method in exactly the same way in which the experimental spectra were analyzed. The chemical-shift differences between the water and the NAA signals in the simulated spectra were then compared with the chemical-shift difference of the known frequency difference, and the deviations were then transformed into temperatures using Eq. [2]. The SDs of temperature values are plotted as a function of each spectral parameter.


Values are given as the mean, SD, or range (median). Comparisons between groups were made, when appropriate, using a paired or unpaired Student's t-test, repeated-measures analysis of variance (ANOVA), or the Statistical Package for the Social Sciences (SPSS, version 10; Chicago, IL, USA).


Eight healthy adult volunteers were studied at an ambient temperature (Tamb) of 20.1–20.5°C (median = 20.2°C; Table 1). During the interval between the start and end of scanning, the ambient temperature was kept constant and there was no significant change in regional body temperature (P > 0.1 for Tt.a, Ttymp, and Toral).

Table 1. Ambient and Regional Body Temperatures With Overall Mean Deep Body Temperatures at Start and End of Scanning*
SubjectAge (years)Sex Temperature (°C)
  • *

    Ambient (Tamb) and regional body temperatures (°C) with overall deep body temperature (°C) values measured at the start and end of scanning. Overall deep body temperature expressed as mean (SD) and calculated from measurements made at Tt.a. (temporal artery), Ttymp, (tympanum), Toral, (sublingual pocket) sites, respectively. M = male, F = female.

146MStart20.137.336.735.936.9 (0.6)6
   End21.036.936.635.936.7 (0.4)7
225MStart21.637.537.236.437.0 (0.5)7
   End20.337.136.736.336.8 (0.3)7
327MStart20.536.636.436.936.5 (0.3)7
   End20.636.336.336.736.4 (0.2)7
452FStart20.536.537.136.436.9 (0.4)8
   End20.736.437.236.336.9 (0.5)10
550MStart20.536.736.336.236.6 (0.2)7
   End21.036.736.736.436.4 (0.5)7
623FStart20.037.637.036.437.1 (0.5)7
   End20.537.437.336.937.3 (0.2)11
726MStart20.036.736.236.236.2 (0.3)7
   End20.436.736.836.836.8 (0.0)7
829FStart20.136.936.836.936.8 (0.2)7
   End20.136.836.736.936.8 (0.2)7

To assess the effect of partial water suppression on the LW and chemical shifts, values were compared from voxel 4 (Fig. 1). The LWs for water for nonsuppressed and 50% suppressed spectra were 5.71 (0.82) Hz and 5.77 (0.54) Hz, respectively, and for NAA they were 4.77 (0.99) Hz for nonsuppressed and 4.74 (0.66) Hz for 50% suppressed spectra. The SNR for NAA in non-water-suppressed spectra (18.2 (4.7)) was significantly lower (P < 0.001) than the SNR for NAA (26.9(6.2)) in partially water-suppressed spectra. The chemical-shift difference between the two peaks (2.664 (0.004) ppm and 2.664 (0.004) ppm equivalent to 341 (0.5) Hz) was almost identical between nonsuppressed and 50% suppressed spectra. The observation that nonsuppressed and partially suppressed 1H MR spectra yielded comparable brain temperatures agrees with the results by Cady et al. (10). Because partially water-suppressed spectra provided a flat baseline and better SNR for NAA, we used the partially suppressed acquisition for the MRS and MRSI data.

A typical set of water and NAA lines from the partially water-suppressed spectra acquired from the four selected voxels is shown in Fig. 2. Mean body temperatures, as measured by independent techniques and calculated from measurements obtained in three separate body regions, ranged from 36.2° to 37.3°C (median = 36.8°C; Table 1). We report the mean brain temperature from the 1H MRS (Table 2); however, it is obvious from Eq. [2] that the actual temperature may vary ±1.3°C due to the calibration curve (12). In brain, the temperature pertaining to the 1H MRS spectra acquired from single voxels (Fig. 2) ranged from 34.9°C to 37.1°C (median 36.5°C; Table 2). There was no evidence of a systematic difference between the regions (P = 0.46, repeated-measures ANOVA). Overall, the mean single-voxel brain temperature was 36.3 (0.3) °C. This single-voxel mean brain temperature was significantly lower (P < 0.01) than (regional) body temperature (36.8 (0.2) °C). The temperature difference, assuming a single representative value from 1H MRS, (ΔTbrain – Tbody) for each subject ranged from −0.8° to −0.1°C (median = −0.5°C; Table 3).

Table 2. Regional and Mean Brain Temperature (°C) in Healthy Volunteers*
SubjectTemperature (°C)
VoxelMean brain (SD)aMean brain (SD)b
  • *

    1H MRS single voxel brain temperature (°C) in four brain regions. For positions of voxels within region of interest refer to Fig. 1 and text.

  • a

    In each subject, mean brain temperature calculated from temperature values obtained from four single voxels.

  • b

    In each subject, mean brain temperature calculated from temperature values acquired from 1H MRSI temperature map from 32 to 40 (median 34) volume elements. All 1H MRSI spectra acquired with 50% water suppression.

136.536.636.736.736.6 (0.1)37.4 (1.8)
235.536.336.537.036.3 (0.6)37.1 (1.4)
335.934.936.036.535.8 (0.6)37.3 (1.9)
436.836.736.336.036.5 (0.4)36.6 (1.1)
535.535.935.836.836.0 (0.5)36.9 (1.6)
636.236.436.536.636.4 (0.1)36.5 (1.1)
736.636.036.036.636.3 (0.4)37.2 (1.4)
837.136.736.836.436.7 (0.3)36.8 (1.0)
Table 3. Brain-Body Temperature Differences (°C) in Healthy Adult Volunteers*
SubjectTemperature (°C)
  • *

    The temperature differences between brain (Tbrain), mean regional body temperature (Tbody), and regional body temperature measured at the temporal artery, tympanum, and sublingual pocket; Tt.a, Ttymp, Toral, respectively. Values for Tbody are the mean of the regional temperature measurements (start and end of scanning measurements).

120.137.336.735.936.8 (0.5)36.6 (0.1)−0.7−0.10.7−0.2
220.337.537.236.436.9 (0.4)36.3 (0.6)−1.2−0.9−0.1−0.6
320.536.636.436.936.5 (0.2)35.8 (0.6)−0.8−0.6−1.1−0.7
420.536.537.136.436.9 (0.4)36.5 (0.4)0.0−0.60.1−0.4
520.536.736.336.236.6 (0.2)36.0 (0.5)−0.7−0.3−0.2−0.6
620.037.637.036.437.2 (0.3)36.4 (0.1)−1.2−0.60.0−0.8
720.036.736.236.236.5 (0.4)36.3 (0.4)−−0.2
820.136.936.836.936.8 (0.2)36.7 (0.3)−0.2−0.1−0.2−0.1

1H MRSI provides a time-efficient means of acquiring spectra from large brain volumes (13). Temperature maps were generated from 1H MRSI data for each volume element throughout the region of interest (ROI; see Fig. 1). Figure 3 shows water and NAA peaks for a volunteer from the set of individual MRSI voxels and the temperature map derived from the chemical-shift difference between these peaks. The average LWs of water and NAA were 12.1(1.0) and 13.1(1.0) Hz, respectively. The SNR was 32.5 (4.2) for NAA. The average brain temperature from the MRSI maps (37.0 (0.3) °C) was significantly higher (P < 0.01) than that derived from the four volumes of single-voxel 1H MRS (36.3 (0.3) °C). However, in MRSI volume elements located close to the single-voxel 4 (close to the center of the FOV), brain temperature (36.7 (0.8) °C), was not significantly different (P = 0.66) from that in the single-voxel volume 4 (36.6 (0.3) °C). The single-voxel 2 temperature (36.2 (0.6) °C) was not different (P > 0.05) from the MRSI map (37.4 (1.7) °C). In the MRSI volume that closely matched single-voxel 3, temperature readings were higher (38.0 (1.6) °C; P < 0.05) than those obtained by single-voxel MRS (36.3 (0.3) °C).

Figure 3.

MRSI of subject 2: (a) water, (b) NAA, and (c) the corresponding temperature map given in the color key (°C). FOV = 230 mm, TR = 1500 ms, TE = 60 ms, SW = 2000 Hz, 16 × 16 MRSI data matrix, 50% water suppression.

The inherent errors of the temperatures for both single-voxel 1H MRS and MRSI methods were estimated using NAA phantom measurements. Because the temperature of the phantom was homogeneous, variation had to arise from measurement and/or analysis factors. Analysis problems may reflect the fact that the baseline, phase, and frequency of the signals correlate with each other, leading to inaccuracy in the line-shape fitting process. The errors in in vivo MRSI data vary spatially, being larger in the outer parts of the brain caused by baseline problems largely due to residual lipid peaks. However, this effect is avoided in the NAA phantom data set. Variation in MRS temperature values in the NAA phantom was ±0.2°C by single-voxel MRS and ±0.8°C by MRSI, which indicates the level of inherent accuracy obtainable under these experimental conditions.

A potential source of inaccuracy in determining temperature by 1H MRS may be spectral characteristics. The simulations on effects of LW (Fig. 4a), SNR (Fig. 4b), and digital resolution (Fig. 4c) on the accuracy of temperature estimation show that the main source of variation appears to be the LW. LW above 8 Hz steeply increased the SD of temperature values. SNR above 15 had a very small effect on the scatter of the NAA-to-water chemical-shift difference. Poor point resolution (>1.5 Hz/point) increased inaccuracy; however, within the range used in the study, the effect appears to be <0.1°C in SD.

Figure 4.

Effects of spectral LW (a), SNR (b), and digital resolution (c) on the SDs of the temperature readings as determined by 1H MRS. 1H MRS spectra with varying spectral characteristics were simulated (see Materials and Methods) and analyzed to determine the NAA-to-water chemical-shift difference. The SDs of the resulting temperature readings are shown for each parameter.


In this study we used the chemical-shift difference between water and the NAA resonance of CH3 at 2.01 ppm to estimate brain temperatures from 1H MR spectra acquired at 3T from single voxels in four different brain regions, at a level just above the lateral ventricles. In addition, we also used MRSI to acquire 1H MR spectra at the same level but from a larger brain volume to generate temperature maps throughout the ROI. Although it is time-efficient to acquire 1H MR spectra from a large brain volume, using MRSI we found that for clinical purposes better accuracy and reproducibility are achieved using single-voxel spectroscopy. From phantom experiments, the temperature accuracy of single-voxel 1H MRS is estimated to be approximately 0.2°C. Although this estimate was obtained at the extreme of the calibration curve, we expect the accuracy to be as good within the physiological range, where the chemical-shift difference between water and NAA at 30–40°C is linearly dependent on temperature. In this regard, we believe that brain temperature values acquired from single voxels provided the most accurate MRS estimation of brain-tissue temperature within the ROI.

The MRS method per se is very accurate for determining the NAA-to-water chemical-shift difference. The main source of inaccuracy in terms of the absolute temperature is calibration (Eq. [2]). Absolute temperature measurement is likely to be less accurate; however, temperature variation can be estimated accurately by MRS, as recently shown by Katz-Brull et al. (14). MRS yields a much better LW compared to 1H MRSI, and our simulation data indicate that the LW is the most likely cause of inaccuracy. Despite the difference between the calibration data used by our group (12) and Katz-Brull et al. (14), the values for absolute brain temperature in close anatomical regions were identical (36.6°C).

Although there is some variation in the temperature difference between the respective brain–body pairs in the healthy adult, in most instances the difference is generally small (see Table 3). This fits well with theoretical modeling of human brain temperature under conditions of normothermia, where any difference between brain and core temperature is expected to be small (2). Our data also agree with observations made in afebrile primates (15), in which a small gradient exists between brain and core temperature (the latter as determined in aortic arch blood). In primates (15) and humans (16), incoming arterial blood exerts a major influence on the thermal environment of the brain, with fluctuations in aortic blood temperature due to feeding, sleeping, and arousal producing parallel changes in brain temperature over time. We have also observed this feature of brain–body temperature relationships in humans, albeit in injured febrile and afebrile brains, without any evidence of “lag” (6).

In healthy adult humans studied at an ambient temperature well below thermoneutrality (∼20°C), we have shown that the magnitude of the brain–body temperature difference appears to be broadly comparable to that reported in studies of healthy rodents (17), primates (15), and humans (18). Although the direction of the temperature gradient between the brain and the body compares with data in healthy rodents, in which brain temperature (measured using 1H MRS) is slightly lower than rectal temperature, this apparently is not the case for primates. In monkeys, the brain (irrespective of anatomical region) was 0.3–0.5°C warmer than the aortic blood (15). In a study of human volunteers (18) the global brain temperature, as assessed from the venous (internal jugular vein) to arterial (aortic blood) temperature difference, was 0.3°C. Thus, in both primates and healthy humans the difference was small but positive, which suggests that the brain has a slightly higher temperature than the body core. It should be noted that our 1H MRS data do not challenge this view.

The differences in the direction of the brain–body temperatures observed in our study compared to those reported by Hayward and Baker (15) and Nybo et al. (18) may be explained in part by the practicalities of obtaining a “true” core temperature (i.e., arterial blood temperature) in healthy subjects. The “ideal” core body temperature measurement is most accurately achieved by invasive monitoring, but this is not practical for volunteers undergoing MR scanning studies. Furthermore, since it is known that the internal organs and deep tissues have slightly different temperatures (19, 20), small variations in the magnitude and direction of brain–body temperature differences between studies might be expected depending on where the body core temperature is measured.

In experimental clinical studies, body temperature measurement is, of necessity, limited to just a few sites. By including in this study of healthy volunteers three noninvasive sites (the tympanum, oral cavity, and temporal artery) in which temperature is commonly measured clinically, we aimed to explore both the direction and magnitude of brain–body temperature differences under normal physiological conditions in order to better understand the differences we previously observed between brain and body temperatures in patients who had experienced brain trauma (6). Even with the limitations inherent in the measurement of “core” temperatures, as used in this study, the differences between the brain and body temperatures of healthy volunteers do not differ greatly in magnitude compared to those of primates or brain-injured patients (6), but the direction does. The problem, for purposes of comparison between species, is that the techniques used for brain temperature measurements differ, as does the site for “core” temperature measurement. With new developments in thermometry and the availability of MR-compatible temperature sensors, the problems inherent in obtaining deep body core temperatures during scanning may be resolved.

Single-voxel 1H MRS of human brain temperature has also provided important information that may aid in the interpretation of brain temperature values obtained using direct, invasive monitoring techniques in neurosurgical patients. It has never been clear whether a single thermistor implanted approximately 3–4 cm into deep WM (usually of the frontal lobe) can reliably predict temperature in other brain regions. Voxel 1 (see Fig. 1 for anatomical reference) illustrates the typical site for placement of intraparenchymal sensors during routine neurointensive care. In the absence of any systematic difference between the focal brain regions within the ROI in healthy subjects, it seems reasonable to assume that measurements obtained via implanted thermistors probably reflect the temperature of other, closely related, undamaged brain regions in patients. In this regard, the present experimental data agree with previous conclusions drawn from theoretical modeling of brain temperature distribution (16).

When brain temperature is measured as a part of neurointensive care (6), the prevailing view (3) is that the temperature of the injured brain will exceed that of the body core, on average by 1–2°C (4, 5). We can confirm that large increases in brain temperature that lead to a positive brain–body temperature gradient of the order of 1–2°C are not a characteristic of human brain–body temperature differences under physiological conditions. By contrast, we (6) and others (21, 22) have observed small differences (both slightly negative and slightly positive) between brain and core temperature after head injury, and it has been difficult to explain why such opposite results are obtained. If we assume that after primary brain damage due to, e.g., contusions and hematoma, the thermal baseline is aortic blood (15), then provided that cerebral perfusion is maintained optimally for adequate brain blood flow, the removal of heat from the brain (especially in the absence of raised cerebral metabolic heat production) will continue without net cerebral heat storage. Any difference between brain and core temperatures would be minimal. This appears to be the case for pharmacologically sedated, ventilated patients (6). On the other hand, cerebral heat retention was observed in primates (15) during a brief interruption of cerebral blood flow (CBF), suggesting a failure of convective heat loss to maintain cerebral heat balance. With cessation of CBF (15), brain temperature fell rapidly, which confirms observations in brain-injured patients following catastrophic reductions in CBF after brain injury (6, 8, 23).

So far, discussion has focused on comparing the differences between brain and body temperature using 1H MRS. Temperature mapping using MRSI techniques is an attractive option because it is a quicker method for achieving large brain coverage and is gaining clinical interest (13). However, our results indicate that single-voxel MRS and MRSI temperature mapping in a clinical 3 T scanner yield consistent values only for the brain area in the center FOV for MRSI. In the periphery, considerable heterogeneity is evident that does not exist across the ROI covered by single-voxel spectroscopy. Therefore, we cannot be confident that the apparent thermal mosaics produced by MRSI temperature maps reveal genuine variations in physiological temperature. The LWs in MRSI are broader than those in single-voxel spectra, even though the former has smaller nominal volume elements than the latter. For these two reasons the SNR is low in MRSI voxels, however, low SNR can not explain entirely the observed inaccuracy of MRSI, because the values in the center of the MRSI FOV match single-voxel MRS temperatures. The inaccuracy in MRSI temperature mapping under present conditions is likely to be due to broad lines (Fig. 4), but it may be further exacerbated by baseline variation together with the relatively small dynamic range of MRS thermometry at 3T (i.e., 1.2 Hz per °C). Thus, despite the potential for faster data acquisition using MRSI in clinical settings, future improvements in MRSI sequence design and MRS hardware are required to facilitate accurate spectral data collection for temperature mapping in large brain volumes. If such improvements can be achieved, brain-temperature mapping by MRSI could be used to investigate human thermoregulation in health and disease, and may potentially aid in diagnosing and monitoring therapy in patients with brain conditions.

A limitation to the use of MRS and MRSI in the setting of brain pathology is the presence of blood breakdown products and the dependence of the technique on NAA concentration per se. In cases of neuronal death (e.g., due to stroke or brain trauma), the concentration of NAA will fall, leading to poor SNR of this temperature reference. This poses an important limitation to the use of noninvasive brain-temperature estimation using the chemical shift of water referenced to NAA. Other 1HMRS peaks, such as creatine and lactate, can be used in place of NAA (24). However, blood breakdown products may broaden the MRS lines and render the method inaccurate for temperature determination.

In conclusion, this study has shown that small differences only exist between brain and body temperatures in healthy subjects under standard physiological conditions. Currently, single-voxel spectroscopy is more accurate than MRSI for obtaining regional brain temperatures. When conventional temperature monitoring sites are used to determine the brain–body temperature gradient in healthy human subjects, the average brain temperature may be slightly lower than the body temperature; however, given the accuracy of the 1H MRS method for determining absolute temperature, the difference may go either way. Subtle differences in the direction of the gradient between humans and primates may be explained by differences in the techniques and measurement sites used.


We thank Barry Whitnall for expert technical assistance. This study was supported in part by the Medical Research Council, UK (fellowship to R.V.).