Temperature mapping considerations in the breast with line scan echo planar spectroscopic imaging

Authors


Abstract

A line-scan echo planar spectroscopic imaging (LSEPSI) sequence was used to serially acquire spectra from 4096 voxels every 6.4 s throughout the breasts of nine female subjects in vivo. Data from the serial acquisitions were analyzed to determine the potential of the technique to characterize temperature changes using either the water frequency alone or the water-methylene frequency difference. Fluctuations of the apparent temperature change under these conditions of no heating were smallest using the water-methylene frequency difference, most probably due to a substantial reduction of motion effects both within and without the imaged plane. The approach offers considerable advantages over other methods for temperature change monitoring in the breast with magnetic resonance but suffers from some limitations, including the unavailability of lipid and water resonances in some voxels as well as a surprisingly large distribution of water-methylene frequency differences, which may preclude absolute temperature measurement. Magn Reson Med 58:1117–1123, 2007. © 2007 Wiley-Liss, Inc.

Most of the magnetic resonance (MR) parameters that contribute to the signal intensity of MR images are temperature sensitive and so of potential utility for guiding thermal therapies (1–5). Of the available MR parameters, there is a growing consensus that the temperature sensitivity of the water proton resonant frequency (PRF) is most suited for guiding thermal therapies. Utilizing the water PRF is advantageous because it appears to be tissue type independent, at least in nonfatty tissues (6). Furthermore, the PRF appears to be minimally affected by heat-induced tissue changes and can be accessed using standard MR methods like gradient echo imaging sequences (7–10). Overall, the PRF appears to be more sensitive than other endogenous parameters like water relaxation times or water diffusion (4). The PRF method has been shown to be useful in numerous animal studies for online guidance of thermal ablation (10–13) and its use has been demonstrated in several clinical studies (14–19). The temperature sensitivity of the water PRF arises from changes in the degree of electron screening induced by temperature-induced variations in the hydrogen bonding network (20).

Typically, changes in water PRF are estimated using phase-difference images acquired with gradient echo sequences (7). Difference images are generally employed to remove the phase variations caused by spatial inhomogeneities in the magnetic field. Because of the use of image subtractions, the method is sensitive to non-temperature-related effects that alter the local magnetic field between scans. These changes can include drift in the static field (21), patient motion outside of the imaging plane (22), and local or global changes in magnetic susceptibility, which can be altered by heating itself (23, 24). The sensitivity of the phase-difference MR thermometry to motion sensitivity is particularly problematic. For example, when motion occurs between scans, phase maps acquired during heating are no longer aligned with the baseline scans used in the phase subtractions and so can result in unacceptably large errors relative temperature estimations, as previously found in the breast (17), with Peters et al. (25) reporting respiration-induced errors of up to 0.2 ppm, which would correlate to a temperature error of 20°C.

In efforts to reduce the artifacts associated with the water phase-difference approach, internal referencing of the water frequency to temperature-insensitive resonances in the proton spectrum has been suggested. In the brain, the proton resonances from low concentration brain metabolites N-acetyl aspartate, choline, and creatine have been suggested for this purpose (26–29). The methylene lipid resonance in breast or liver (30, 31) and the citrate proton signal in the prostate (32) have also been suggested as internal temperature reference signals. Single-voxel spectroscopic approaches (26–28) and spectroscopic imaging methods (28, 31, 33, 34) have been employed for such studies, with subsequent signal-to-noise and scan time considerations becoming issues for any specific application.

In this work, a rapid line-scan echo-planar spectroscopic imaging (LSEPSI) sequence (35, 36) previously demonstrated for temperature monitoring (34) was tested for potential temperature monitoring in the breast. Specifically, data from serial LSEPSI acquisitions were analyzed to determine the uncertainty in temperature difference measurements based on the water/lipid frequency difference over the course of several minutes and compared with the uncertainty in temperature measurements based on the water frequency alone. In addition, the percentage of typical breast voxels that can yield spectra suitable for rapid fat/water frequency measurement with LSEPSI was evaluated as was the potential to perform absolute temperature measurements throughout the breast.

MATERIALS AND METHODS

Our institutional review board approved the imaging protocol and all volunteers provided written informed consent. Studies were performed with a 1.5 T scanner (General Electric Medical Systems, Milwaukee, WI, USA). Nine women, age range 26–67 years with median age of 51 years, participated in the study. Subjects were positioned feet first in the prone position in the scanner with a 12-cm-diameter surface receive only coil (USA Instruments, Aurora, OH, USA) placed under one breast. The body coil was used for RF transmission. Fat-suppressed fast spin-echo axial T2-weighted images with 3-mm slice thickness and repetition time/effective echo times (TR/TE) of 2000 ms/80 ms were acquired with a 16-cm FOV to identify a single slice for spectroscopic interrogation. Spectroscopic imaging was then performed with the LSEPSI sequence (34–36). Briefly, LSEPSI uses a 90°–180° pair of RF pulses to select orthogonal slices inclined by 45° from the normal of the plane to be imaged. A spin-echo from the column defined by the intersection of the slices forms at an TE of 10 ms and the second half of this echo is read out with 32 asymmetric gradient echoes spaced by 3.1 ms to “simultaneously” encode along column spatial and spectral dimensions (36) with a spectral resolution of 10 Hz. A receiver bandwidth of ± 48 kHz with 128 frequency-encoding steps per gradient echo was employed with a 16-cm FOV so that the along-column spatial resolution was 1.25 mm. A total of 32 adjacent columns were sequentially sampled at 200-ms intervals to cover the second in-plane dimension for a final 16 × 16 cm2 spatial FOV in a total scan time of 6.4 s. The columns were acquired such that the image plane was swept in two passes, even columns in the first pass and odd columns in the second, to avoid T1 saturation. We utilized a slice selection thickness of 4 mm with resulting nominal voxel volumes of 0.025 ml. In three subjects, additional LSEPSI sequences were performed using 8 mm slice thicknesses, a 32-cm FOV and a 2.5-ms gradient echo spacing with a spectral resolution of 12.5 Hz and a nominal voxel volume of 0.05 ml.

For each breast, a time series of 50 sequential LSEPSI acquisitions was acquired at 6.4-s intervals for approximately 5.3 min. The overall scan time was approximately 45 min per subject including setup and prescans. Spectroscopic image reconstruction and subsequent data analyses were performed with Matlab software (version 6.1; The Mathworks, Natick, MA, USA). Reconstruction was performed with a two-dimensional Fourier transformation (2D-FT) after zero-filling of the spectral dimension to 128. Magnitude spectra were analyzed by fitting the water and methylene resonances with Gaussian functions using the Levenberg-Marquardt nonlinear least squares algorithm to determine central frequencies and areas. For very fatty voxels in which the water peak was poorly represented, the resonance from the olefinic lipid protons around 5.3 ppm was fit instead of the water peak. This data was included here to confirm that the broad range of water-methylene difference frequencies observed in the volunteers (see below) was not due to an error in the fitting routine. A narrow distribution of methylene-olefinic frequencies would be expected. Finding such a distribution in our data would suggest that errors were not occurring with our fitting routine. The spectral fits were first performed on the highest signal-to-noise spectroscopic image available from the complete data set as generated by signal averaging all 50 acquisitions per breast. Next, spectral fitting was performed on the spectroscopic images acquired at each individual time point. Voxels with poorly fit spectra were excluded by visual inspection and the total percentage of voxels within each breast where Gaussian fits were deemed adequate for both water and methylene peaks was calculated. Previous study using this sequence found that despite its low spectral resolution, after the Gaussian fit, we could track small frequency shifts with this pulse sequence (34). Fits to a Gaussian instead of a Lorentzian function were used because the Gaussian fits were found to better match the magnitude spectra, although both should yield similar values for the resonant frequency. A homogeneous water/lipid phantom consisting of equal volumes of mayonnaise and lemon juice held at a constant temperature of 18°C was subject to the experiments and analyses used for the in vivo breast studies in order to determine basic temperature stability properties in the absence of motion.

Estimates of temperature change stability over the 5.3-min LSEPSI acquisitions were based on analyses of the fluctuations of the water frequency and from the water and methylene frequency difference. To do so, for each voxel with adequately fit spectral peaks, the water frequency fw or the water-methylene fw-m frequency difference at each time point was subtracted from the baseline value acquired in the first acquisition. The resulting frequency differences, which would vanish in the limit of no frequency fluctuations, were converted to apparent temperature changes using Hindman's value of –0.01 ppm/°C (20). Means and standard deviations (SDs) of the observed temperature changes were calculated from the available voxels at each time point and were plotted as a function of time. To characterize the magnitude of the apparent temperature fluctuations, the root mean square (RMS) of the mean apparent temperature fluctuations averaged over all 50 time points was calculated for the water frequency–only approach, RMS(fw), and for the water-methylene frequency difference approach, RMS(fw-m). For each subject, a paired t-test was performed to determine whether there was a statistically significant difference between the RMS values calculated with the water frequency alone vs. the water-methylene frequency difference.

In addition to the stability tests, the distribution of the water-methylene frequency difference throughout the breasts and the phantom was examined from the 50 time point–averaged spectroscopic images. Assuming that this distribution is due only to spatial variations of absolute temperature across the breast, variations on the order of 2°C with a rather small distribution of fw-m of less than 5 Hz would be predicted.

RESULTS

Both the water and the methylene peaks were detected in many locations in the breasts, although in many locations only one or the other peak was quantifiable with the Gaussian fitting approach. The percentage of voxels within the breast where both methylene and water resonances were quantifiable ranged from 27% to 60% in the averaged images and 3% to 34% at the single time-points for the 4-mm-thick slice acquisitions (Table 1). Predictably, a much larger percentage of voxels in the breast had quantifiable methylene and water resonances when 8-mm-thick slices were employed, as shown for the three women for whom both data sets were acquired (Table 1). In the phantom experiment, Gaussian fits to both resonances was possible in all of the spectra from the 50 time point–averaged data while diminished signal-to-noise in the single time point data reduced the ability to fit approximately 45% of the spectra with double Gaussian functions. Example breast spectra and spectroscopic images are shown in Figs. 1 and 2, respectively.

Table 1. Percentage of the Breast and Phantom Where Both the Methylene Lipid Peak and the Water Peak Were Detected for the Nine Volunteers and the Water/Lipid Phantom, Respectively
SubjectSlice thickness
4 mm8 mm
50 NEX (%)1 NEX (%)50 NEX (%)1 NEX (%)
  1. NEX = number of excitations.

14518-29
22912-17
33010-19
45526-347747-74
53514-215940-53
64720-306947-62
76027-38
82713-20
9333-18
Mean4016-256845-63
Phantom10053-60
Figure 1.

Magnitude spectra and double Gaussian fits from two different locations in a volunteer breast. The spectra were aligned in this figure so that the methylene peaks were at −1.33 ppm. In the top example, the water and methylene peaks were evident. In the bottom example, the water peak was too small to detect reliably and the small olefinic peak was fit instead. These examples are from data averaged over 50 acquisitions with a slice thickness of 4 mm.

Figure 2.

a,b: Maps of the lipid (a) and water magnitudes (b) in the breasts of a volunteer. c: Image of the distribution of water-methylene frequency difference. This frequency difference was generally lower near the skin and the chest wall in this case. This could not be explained by a temperature distribution across the breast, since a lower frequency difference correlates to a higher relative temperature. d: Map showing where both water and methylene peaks were detected and where the methylene and olefinic lipid peaks were fit instead. These examples are from data from one subject averaged over 50 acquisitions with a slice thickness of 4 mm.

The water-methylene frequency difference was more stable over time than the water frequency alone, as shown by the plots of the apparent temperature change as a function of time in Fig. 3 for both 4-mm- and 8-mm-thick slice acquisitions of the left and right breasts in one subject, as calculated by averaging all available voxels at each time point to obtain the means and SDs shown. Fluctuations of the mean temperature change were as large as 5°C when the water frequency alone was used while they were generally less than 1°C with the frequency difference approach. The use of 8-mm- vs. 4-mm-thick acquisitions had little effect on the apparent temperature change fluctuations with time though use of the thicker slice generally resulted in smaller SDs of the apparent temperature fluctuations due to the improved signal-to-noise. The spread of the apparent temperature change fluctuations, as gauged by the mean RMS values, was on the order of 2°C for the water frequency alone but less than 1°C when the water-methylene frequency difference was employed, as shown in Fig. 4 for the 4-mm-thick slice data of all nine subjects and also for the phantom. In eight of the nine volunteers the difference between the fw-based and the fw-m-based RMS values was statistically significant (P < 0.01), as it was also for the phantom (Fig. 4).

Figure 3.

Plots of the apparent temperature change as a function of time for the water peak alone and for the water-methylene peak frequency difference in one volunteer for 4-mm and 8-mm slice thickness acquisitions for both the left and right breasts as calculated from averaging over all available voxels at each time point to get a mean and SD. Substantial variation with time was observed in the breast when the temperature variation was measured with the water peak alone in the breast. This variation was considerably reduced when temperature variation was measured with the water-methylene frequency difference.

Figure 4.

Root mean square (RMS) (mean ± SD) of the apparent temperature variations for the nine volunteers and for the water/lipid phantom as found from using the water peak alone and with the water-methylene frequency difference. In eight of nine volunteers and even in the phantom, the RMS was significantly less (P < 0.01) when measured with the water-methylene frequency difference than with the water peak alone.

Figure 5 shows histograms of the water-lipid difference frequencies in one volunteer (Fig. 5, left) and in the phantom (Fig. 5, right), as measured from the fully averaged data sets of each. The distribution of the water-methylene frequency difference was significantly broader in the volunteers than in the phantom. The distribution of the olefinic-methylene frequency difference, however, was similar to the water/methylene distribution in the phantom. Table 2 summarizes the water-methylene frequency distributions for all of the subjects. The SD of the distribution of fw-m corresponded to temperature distributions on the order of ±14°C. The corresponding value in the mayonnaise/lemon juice phantom was only ±2°C. Plots of fw-m vs. relative peak areas as represented by water/(fat + water) fractions revealed no correlation between the water-methylene frequency difference and the relative peak areas in any of the volunteers.

Figure 5.

Distribution of the methylene-water and methylene-olefinic difference frequencies for images acquired in a volunteer (left) and in a water/lipid phantom (right).

Table 2. Mean Frequency Difference Between the Methylene Lipid and Water Peaks (w-m) and Between the Olefinic Lipid and Water Peaks (o-m) for the Nine Volunteers and for the Water/Lipid Phantom (at 18°C)
SubjectSlice thickness
4 mm8 mm
Mean (w-m) (ppm)± (ppm)Mean (o-m) (ppm)± (ppm)Mean (w-m) (ppm)± (ppm)Mean (o-m) (ppm)± (ppm)
1−3.310.12−3.990.03
2−3.350.14−3.980.03
3−3.340.16−3.980.03
4−3.260.08−4.000.02−3.270.09−4.000.03
5−3.320.12−4.000.06−3.290.10−3.990.02
6−3.310.11−4.010.07−3.310.10−3.980.02
7−3.380.14−3.970.03
8−3.350.16−3.980.03
9−3.310.19−4.110.13
Mean−3.320.14−4.000.05−3.290.10−3.990.02
Phantom−3.590.02

Figure 6 shows a stack of breast spectra from voxels extracted from the 50 time point–averaged data set of one subject. The spectra have been ranked from bottom to top in this plot according to a decreasing water-methylene frequency difference fw-m and have been arranged with all the methylene peaks manually aligned at –1.3 ppm. The olefinic peak is visible as a faint peak to the left of the water peak in Fig. 6. The shift of the water frequency with respect to the methylene frequency may be visualized in Fig. 6, where a solid line running vertically from bottom to top, first through the olefinic resonance and then through the water resonance, has been drawn.

Figure 6.

Spectra from all of the voxels in the breast of a volunteer. The spectra were sorted by the size of the water-methylene and or olefinic-methylene frequency difference and are arranged so that the methylene lipid peaks are aligned at –1.3 ppm, as indicated by the vertical black line on the right side of the figure. The vertical black line on the left side of the figure passes through the center of the olefinic resonance in the lower portion of the plot and transitions to the water resonance once this peak became quantifiable with the Gaussian fit. The water frequency is observed to vary toward the methylene resonance even as the methylene-olefinic frequency difference remains unchanged.

DISCUSSION

The primary finding of this work is that fluctuations associated with temperature change measurements in the breast can be considerably reduced by using the frequency difference between water and methylene signals as opposed to the water frequency alone, most probably due to partial or entire elimination of motion effects both within and outside of the imaging plane. The approach can be used for many but not all voxels within typical breasts though signal averaging and/or decreasing spatial resolution can be used to increase the number of voxels suitable for the frequency difference approach, at the expense of increased scan time and/or partial volume effects. Alternatively, in water-dominated voxels where a lipid resonance is difficult to quantify it might be feasible to use a reference lipid signal from a neighboring voxel, assuming that motion or susceptibility related changes have relatively smooth spatial variations which accommodate interpolation.

The LSEPSI sequence has several advantages over other methods proposed for using the lipid signal as a reference. One can tailor the sequence to a particular situation by obtaining lines only in the area of interest, which could improve the temporal resolution or allow for additional signal averaging. The line-scan acquisition also reduces T1-saturation compared to short TR 2D-FT EPSI schemes, though it is much less efficient from a volume coverage perspective. The snapshot nature of the individual column acquisitions also provides a relative insensitivity to in-plane motion that would prove more problematic with 2D-FT planar acquisitions (37). This technique might be further improved by combining it with other motion correction schemes that have been proposed, such as the use of navigator echoes (38), gating schemes (39), and removal of susceptibility-induced magnetic field variations through extrapolation of phase data in surrounding (nonheated) regions (40), since it uses data that is colocalized with the heated zone that suffers exactly the same artifacts.

Somewhat surprisingly, a relatively large distribution of water-methylene difference frequencies was observed in the breasts that was not observed in the mayonnaise/lemon juice phantom. The findings are consistent with an initial report by our group that utilized fewer signal averages and a less standardized set of LSEPSI parameters (41). Temperature variations throughout the breast may partially explain some of the water-methylene frequency distribution, but clearly not all of it. The average SD of the frequency difference was ±0.14 ppm, which would correspond to ±14°C. In addition, the largest difference frequencies, which would correspond to the coolest areas, were often located in central locations of the breast, inconsistent with likely temperature distributions. These results thus indicate that absolute temperature measurements in the breast using the water-methylene frequency difference are probably not feasible. The reason for this relatively large distribution of water-methylene difference frequencies remains unclear. Several other factors could potentially affect the difference frequency, such as the pH and/or protein content (26). Alternatively, susceptibility differences of fat and water as inhomogeneously mixed throughout the breast, unlike the well-mixed phantom material, may play a role. Each of these hypotheses is speculative and would require further experimental and/or theoretical substantiation. Significantly, the distribution of the lipid-olefinic frequency difference was small, indicating that the large water-methylene frequency distribution was not an artifact from any systematic error associated with the acquisitions or analyses but was definitely due to a shift of the water resonance with respect to the methylene signal.

Ancillary