SEARCH

SEARCH BY CITATION

Keywords:

  • chemical-shift imaging;
  • GRASE;
  • Dixon;
  • lipid-water decomposition

Abstract

  1. Top of page
  2. Abstract
  3. TECHNIQUE
  4. MATERIALS AND METHODS
  5. RESULTS
  6. DISCUSSION
  7. CONCLUSIONS
  8. Acknowledgements
  9. REFERENCES

Three-point Dixon techniques achieve good lipid-water separation by estimating the phase due to field inhomogeneities. Recently it was demonstrated that the combination of an iterative algorithm (iterative decomposition of water and fat with echo asymmetry and least-squares estimation (IDEAL)) with a fast spin-echo (FSE) three-point Dixon method yielded robust lipid-water decomposition. As an alternative to FSE, the gradient- and spin-echo (GRASE) technique has been developed for efficient data collection. In this work we present a method for lipid-water separation by combining IDEAL with the GRASE technique. An approach to correct for errors in the lipid-water decomposition caused by phase distortions due to the switching of the readout gradient polarities inherent to GRASE is presented. The IDEAL-GRASE technique is demonstrated in phantoms and in vivo for various applications, including pelvic, musculoskeletal, and (breath-hold) cardiac imaging. Magn Reson Med 57:1047–1057, 2007. © 2007 Wiley-Liss, Inc.

Successful decomposition of the water and lipid components is an important problem in MRI. In many clinical applications the bright lipid signal may obscure relevant anatomical features, in which case good lipid suppression is needed. In other cases, detecting the presence of lipid is important for diagnostic purposes. Some examples are the diagnosis of fatty livers (1), the characterization of certain types of neoplasms (i.e., lipomas, adrenal adenomas, and renal cell carcinomas) (2–5), and the diagnosis of arrhythmogenic right ventricular dysplasia, which is characterized by lipid infiltration in the right ventricular free wall (6–8).

In current clinical practice, lipid-water separation is mainly performed with chemical-shift selective saturation (9–13) or inversion recovery (IR) methods (14–17). Chemical-shift selective saturation is susceptible to field inhomogeneities and typically yields nonuniform suppression over the field of view (FOV). IR methods result in longer scan times, and partial cancellation of the chemical-shift component is observed due to partial recovery after inversion. As an alternative, phase-sensitive chemical-shift imaging has been an active area of research since the method was first proposed by Dixon in 1984 (18). These methods are based on collecting multiple data sets with different sequence timing to control the phase difference between chemical shift components.

Originally proposed as the two-point Dixon method (18), the three-point Dixon method or its variations have been further developed as a means to generate lipid and water images without the effects of field inhomogeneities (19–27). However, the success of these methods to correct for field inhomogeneities relies on phase unwrapping or other phase-estimation techniques that are sensitive to global tissue connectivity, artifacts, and noise in the data.

In three-point Dixon implementations with spin-echo (SE) or fast spin-echo (FSE) methods, the acquisition of the middle phased-shifted point at the SE point (which implies symmetric data acquisition) does not generate robust lipid-water separation when the lipid : water ratio is ∼1 (28). This has been shown by calculation of the theoretical maximum effective number of signal averages (NSA) (29) and other analytical techniques (23). Better lipid-water separation can be achieved when the middle point is shifted from the SE point so that the phase shift between lipid and water is π/2 + kπ, where k is an arbitrary integer (23, 28, 29).

The combination of an iterative lipid-water decomposition algorithm (30) and a three-echo data acquisition with the center echo shifted relative to the SE point (iterative decomposition of water and fat with echo asymmetry and least-squares estimation (IDEAL)) is reported to generate a robust and accurate separation of water and lipid (28). IDEAL has been implemented with FSE, gradient-echo (GRE), and steady-state free precession (SSFP) sequences, and has been demonstrated in body (28, 30), musculoskeletal (28, 30), and cardiac imaging (31). IDEAL requires the acquisition of at least three images with different phase shifts between lipid and water in order to resolve the lipid and water images and the field map. In FSE applications, the collection of the three phase-shifted echoes is done in separate TR periods, with the drawback of relatively long scan times (28).

The use of the gradient- and spin-echo (GRASE) technique is an alternative to FSE acquisitions. GRASE provides efficient data collection (32) as well as the possibility of generating images with various phase shifts within each SE period (33). In this work we present a robust and time-efficient lipid and water decomposition technique that combines a variant of the GRASE method and the IDEAL algorithm. A method to correct for phase errors caused by eddy currents due to the fast switching of the readout gradient polarities inherent to GRASE is also described. The IDEAL-GRASE technique is evaluated in phantoms and in vivo for a series of applications, including pelvic, cardiac, and musculoskeletal imaging.

TECHNIQUE

  1. Top of page
  2. Abstract
  3. TECHNIQUE
  4. MATERIALS AND METHODS
  5. RESULTS
  6. DISCUSSION
  7. CONCLUSIONS
  8. Acknowledgements
  9. REFERENCES

A diagram of the pulse sequence used in IDEAL-GRASE is shown in Fig. 1. The sequence consists of a 90° RF pulse followed by a train of 180° refocusing RF pulses generating a series of SE periods. Within each SE period, several echoes (referred to as En, where n = 1, 2, …, N represents the nth echo in each SE period) corresponding to the same k-space line are collected back and forth by shifting the polarities of the readout gradients. In typical GRASE acquisitions, one SE and multiple GREs (on both sides of the SE point) are acquired. However, as discussed above, the acquisition of data at the SE point is not optimal for lipid-water separation. Thus, in our implementation the three echoes (minimum number of phase-shifted echoes that is required for IDEAL) are shifted relative to the SE point. This leaves extra space within each SE period that can be used for the acquisition of an additional echo, as indicated by the broken rectangular box. If the echo shifts are (–π/6, π/2, 7π/6), which is the combination for a three-echo acquisition scheme reported to yield optimal NSA in FSE acquisitions (28, 29), we need to choose a bandwidth (BW) of ±125 kHz assuming 256 readout points. This echo-shift combination leaves sufficient time to collect a fourth echo at –5π/6. If data are acquired with a BW of ±62.5 kHz with echo shifts (–π/2, π/2, 3π/2), an additional fourth echo can be acquired at –3π/2.

thumbnail image

Figure 1. Schematic diagram of IDEAL-GRASE. For clarity, the slice-selection gradient waveforms are not shown.

Download figure to PowerPoint

In the GRASE sequence reported here, there is an additional SE period after the regular data acquisition period. The data acquired in this SE period are used to correct phase errors caused by switching the readout gradient polarities. As illustrated in Fig. 1, in odd TR periods the polarities of the readout gradients in the phase-correction SE period are the same as those in the regular data-acquisition periods. In even TR periods the polarities of the readout gradients in the phase-correction SE period are flipped relative to the readout gradients in the regular data-acquisition periods. During the phase-correction SE period, the phase-encoding gradients are turned off in order to take advantage of the high SNR for data collected at the center of k-space, and to improve the accuracy of the estimated phase error (34).

Prior to applying the IDEAL algorithm, data acquired in the phase-correction SE period (hereafter referred to as “the calibration data”) are processed in the following way:

  • 1
    For the nth echo, En, the order of all k-space calibration lines acquired with negative readout gradient polarity is reversed to reflect the flipped k-space scan direction.
  • 2
    The Fourier transform of the k-space calibration line of En acquired in an odd TR period is divided by the Fourier transform of the calibration line of this echo acquired in the following (even) TR period. The phase error (Φmath image (x), where x indicates the frequency-encoding direction in the image domain) is calculated as the phase of the quotient.
  • 3
    Step 2 is repeated for all odd-even TR pairs, and Φmath image (x) is averaged to reduce noise in the estimation.
  • 4
    The averaged Φmath image (x) curve is smoothed and then fit to a polynomial. In order to exclude the background noise region in the curve-fitting algorithm, all calibration lines of En collected during the odd TR periods are averaged and then Fourier transformed. The magnitude of the Fourier transform is then thresholded to produce a mask that is used to segment out the background noise region.
  • 5
    The estimated phase error, Φˆmath image (x), obtained from the polynomial fit is halved and subtracted from the image of En in the frequency-encoding direction. This is equivalent to eliminating the phase error in data acquired with positive (or negative) polarity and making it consistent with data acquired with pseudo-neutral polarity. The rationale to use the pseudo-neutral polarity point is to remove the effect of the echo shift on the phase error (further discussion of this concept is presented in the Results section).
  • 6
    Steps 1–5 are repeated for all N echoes and all receivers (for multicoil acquisitions).

MATERIALS AND METHODS

  1. Top of page
  2. Abstract
  3. TECHNIQUE
  4. MATERIALS AND METHODS
  5. RESULTS
  6. DISCUSSION
  7. CONCLUSIONS
  8. Acknowledgements
  9. REFERENCES

The IDEAL-GRASE sequence was implemented on a 1.5T MRI scanner (GE Signa NV-CV/i; General Electric, Milwaukee, WI, USA) equipped with 40 mT/m gradients. All algorithms were implemented in MATLAB (The Mathworks, MA, USA). Informed consent was obtained from patients and volunteers prior to imaging in compliance with institutional review board requirements.

MRI Experiments

A phantom consisting of two vials (diameter = 2 cm) of oil surrounded by water in a 7-cm-diameter container was used to demonstrate the technique.

All phantom and in vivo data with the IDEAL-GRASE sequence were acquired with one of the two following sets of parameters: 1) echo shifts = (–5π/6, –π/6, π/2, 7π/6), BW = ±125 kHz, ETL = 10, NEX = 1, and matrix = 256 × 240 for each echo (excluding the correction data); and 2) echo shifts = (–3π/2, –π/2, π/2, 3π/2), BW = ±62.5 kHz, ETL = 8, NEX = 1, and matrix = 256 × 192 for each echo. Other parameters used to acquire phantom data were FOV = 20 × 20 cm2, TR = 1 s, slice thickness = 6 mm, TEeff = 35 ms with (–5π/6, –π/6, π/2, 7π/6), or TEeff = 42 ms with (–3π/2, –π/2, π/2, 3π/2). In vivo data of the pelvis were acquired with FOV = 44 × 44 cm2, TR = 1 s, slice thickness = 6 mm, TEeff = 35 ms with (–5π/6, –π/6, π/2, 7π/6) and 43 ms for (–3π/2, –π/2, π/2, 3π/2). In vivo knee data were acquired with FOV = 20 × 20 cm2, slice thickness = 4 mm, TR = 1 s, TEeff = 42 ms for (–5π/6, –π/6, π/2, 7π/6). The scan parameters used in cardiac imaging were FOV = 36 × 36 cm2, slice thickness = 6 mm, TR = 1RR, and TEeff = 30 ms for (–3π/2, –π/2, π/2, 3π/2). Phantom and in vivo data were acquired with either a four-element phased array receive-only torso RF coil or a transmit/receive extremity RF coil designed for knee imaging with an extension to image the foot and ankle.

For comparison purposes, data were also acquired with an FSE technique using IDEAL (28) or with the conventional chemical-shift suppression method. IDEAL-FSE data were collected using the same parameters used to acquire the corresponding IDEAL-GRASE data set. For data acquired with FSE using chemical-shift suppression, all parameters except for ETL = 12 were the same as for the corresponding IDEAL-GRASE data set. A larger ETL was used to keep the data acquisition period during the echo train comparable to that in IDEAL-GRASE. In this manner the signal decay due to T2 was similar for both methods.

Water and lipid images were generated using the IDEAL algorithm described in Ref.28. For the in vivo data, a region-growing algorithm was integrated into the IDEAL algorithm to achieve better lipid-water separation. The region-growing algorithm helps prevent ambiguity between water and lipid by utilizing neighboring pixels to provide an improved initial guess of the field map. It also improves the stability of the algorithm by globally searching for a seed pixel (35).

Noise Performance Calculations

Noise performance calculations were based on the signal model:

  • equation image(1)

In Eq. [1] ρw, ρf, ϕw and ϕf are the magnitude and phase of the water and lipid components, respectively; Δϕ is the chemical shift; Δψ is the local magnetic resonance offset; and tn is the time shift for the nth echo from the SE point. Theoretical calculations were performed using the Cramér-Rao bound analysis (29) by estimating the lower bound of the variance in the lipid or water image, σmath image, relative to the variance of a single source image, σmath image. These quantities were then used to estimate

  • equation image(2)

Following the procedure described in Ref.29, the noise performance was also estimated using Monte Carlo simulations. The parameters used in the simulations were ϕw = ϕf = π/4, Δψ = π/20 radians/ms, SNR = 200 for each source image, and 500 independent Gaussian noise realizations for every ρw : ρf ratio (from 0.01 to 100). The variance of the source image and the decomposed water or lipid image obtained with the IDEAL algorithm were used to evaluate the NSA for each ρw : ρf ratio.

RESULTS

  1. Top of page
  2. Abstract
  3. TECHNIQUE
  4. MATERIALS AND METHODS
  5. RESULTS
  6. DISCUSSION
  7. CONCLUSIONS
  8. Acknowledgements
  9. REFERENCES

Phase Correction in IDEAL-GRASE

In order to demonstrate the phase error in GRASE images due to the switching of the readout gradient polarities, we acquired data from a phantom and in vivo. The phantom data were acquired using the extremity coil. In vivo data of the pelvis were acquired using the four-element phased-array coil. Shown in Fig. 2a is the phase error in the phantom for each of the four echoes. Figure 2b shows the phase error for the pelvic data from one of the receivers in the phased-array coil. Although the phase error is nearly linear across the small (7-cm-diameter) phantom, in general the phase error is nonlinear, as shown in Fig. 2b. In both cases the phase error can be well fit by a polynomial, as indicated by the solid lines in the figure. If the phase error were independent of the echo number (i.e., En), it would be sufficient to simply subtract the phase error from the echoes collected with either a single positive or negative readout gradient. However, as shown in the plots, there is a slight dependence of the phase error on echo number. In the proposed algorithm this dependence is removed by halving the estimated phase error from each echo collected in the phase correction SE period prior to subtracting it from the corresponding source image.

thumbnail image

Figure 2. Phase error between data collected with opposite polarities of the readout gradients. a: Phantom data acquired with a transmit/receive extremity RF coil. b: In vivo (pelvic) data acquired with a four-element phased-array RF coil (data shown are from one of the four receivers). The solid curve is the polynomial fit to the phase error (symbol). The background region outside the object is not shown. The axis, x, represents pixels in the image along the frequency-encoding direction. The phase error of odd echoes is obtained by dividing data collected with positive polarity by data collected with negative polarity. The phase error of even echoes is obtained by dividing data collected with negative polarity by data collected with positive polarity.

Download figure to PowerPoint

The effect of the phase error in the lipid-water decomposition from IDEAL-GRASE data is shown in Fig. 3 for the lipid-water phantom. Figure 3a–f show data acquired with the four-element phased-array coil. Figure 3g–l show data acquired with the extremity coil. Note that the lipid-water separation with IDEAL-GRASE without phase correction (Fig. 3a, b, g, and h) is poor (i.e., there is significant mismapping of lipid and water signals) and the error is nonuniform along the frequency-encoding direction (i.e., the horizontal direction). The error in lipid-water separation is more pronounced in data acquired with the extremity coil (Fig. 3g and h). When the phase error correction is applied, a uniform and complete decomposition of lipid and water is obtained (Fig. 3c, d, i, and j) regardless of the coil. The results obtained from IDEAL-GRASE with the phase error correction scheme are comparable to the results obtained by combining the IDEAL algorithm with an FSE method (Fig. 3e, f, k, and l), which requires a separate TR for each echo shift, resulting in longer scan times.

thumbnail image

Figure 3. ad: Decomposed water and lipid images of the phantom using IDEAL-GRASE from data acquired with a four-element phased-array coil: (a) water and (b) lipid images obtained without phase correction, and (c) water and (d) lipid images obtained with phase correction. For comparison the (e) water and (f) lipid images generated using IDEAL-FSE are included. gj: Decomposed water and lipid images of the phantom using IDEAL-GRASE from data acquired with a transmit/receive extremity coil: (g) water and (h) lipid images obtained without phase correction, and (i) water and (j) lipid images obtained with phase correction. For comparison the (k) water and (l) lipid images generated using IDEAL-FSE are included.

Download figure to PowerPoint

Figure 4 shows the effect of the phase error and the phase correction in the lipid-water decomposition for in vivo pelvic data (only the water image is shown). The water image obtained without phase correction (Fig. 4a) has significant residual errors, especially the areas on the left- and right-hand sides of the pelvis (marked with asterisks in the figure). These correspond to areas of greater phase distortion, as can be verified in Fig. 2b. If one accounts for the areas of signal intensity change due to the nonuniformity of the phased-array coil, it can be seen that the error in the lipid-water separation is nonuniform principally in the frequency-encoding (horizontal) direction. As shown in Fig. 4b, lipid-water separation is improved with phase correction and is comparable to the lipid-water separation obtained with the IDEAL-FSE technique (Fig. 4c), with the advantage that data can be acquired three times faster.

thumbnail image

Figure 4. Water images of the pelvis obtained with IDEAL-GRASE (a) without and (b) with phase correction. The asterisks in (a) highlight areas with incomplete lipid suppression. Data were acquired with a four-element phased-array coil with FOV = 44 × 44 cm2, BW = ±125 kHz, echo shift = (–5π/6, –π/6, π/2, 7π/6), ETL = 10, matrix = 256 × 240 for each echo, TR = 1 s, and TEeff = 35 ms, scan time = 24 s. c: Water image obtained with IDEAL-FSE.

Download figure to PowerPoint

Noise Performance in IDEAL-GRASE

As previously reported by Pineda et al. (29), the phase difference between lipid and water spins determines the theoretical noise performance of the water-lipid separation (i.e., the NSA). Figure 5 shows the magnitude NSA of water and lipid data with water : lipid ratios ranging from 0.01 to 100 obtained theoretically (solid line) and with Monte Carlo simulations (symbols). Note that the three-echo acquisition with echo-shift combinations (–π/2, π/2, 3π/2) or (–π/6, π/2, 7π/6) gives uniform noise performance for all water : lipid ratios. As already reported (29), the optimal noise performance for a three-echo acquisition (NSA = 3) is achieved with the echo-shift combination of (–π/6, π/2, 7π/6). To investigate the noise performance of the four echo-shift acquisitions used in IDEAL-GRASE, we performed theoretical and Monte Carlo simulations for the echo-shift combinations of (–3π/2, –π/2, π/2, 3π/2) (data acquired with BW = ±62.5 kHz) and (–5π/6, –π/6, π/2, 7π/6) (data acquired with BW = ±125 kHz). As expected, when a fourth echo is added, the maximum noise performance approaches 4 (see Fig. 5). The noise performance for the four echo-shift combinations of (–3π/2, –π/2, π/2, 3π/2) and (–5π/6, –π/6, π/2, 7π/6) is less uniform across all water : lipid ratios compared to the corresponding three echo-shift combinations. However, the NSA for any water : lipid ratio is greater than that for the three-echo acquisitions.

thumbnail image

Figure 5. Theoretical magnitude NSA (solid line) and Monte Carlo simulations (symbols) of water (top) and lipid (bottom) for various water : lipid ratios and echo-shift combinations.

Download figure to PowerPoint

Because the echo shifts in GRASE are dependent on the receiver BW and number of readout points, we investigated the overall noise performance, including the constraints imposed by these parameters. Thus we define the SNR factor term (SNRfactor) as:

  • equation image(3)

In order to compare the SNRfactor for the echo shifts investigated, we fixed the time to play the echo train to be around 110 ms and the number of TR periods to 24. This keeps the effects of T2 decay constant and the total scan time fixed. With these parameters and assuming 256 readout points, the SNRfactor for the echo-shift combination (–3π/2, –π/2, π/2, 3π/2) was calculated using BW = ±62.5 kHz and ETL = 8. The SNRfactor for the echo-shift combination (–5π/6, –π/6, π/2, 7π/6) was calculated using BW = ±125 kHz and ETL = 10. Note that the number of phase-encoding steps is different for these two cases (192 and 240, respectively); however, since in the calculation we assume that images are reconstructed into a 256 × 256 matrix, in the SNRfactor calculation there is no difference due to voxel size. NSA was obtained from the theoretical simulations shown above. Plots of SNRfactor are shown in Fig. 6 for (a) water and (b) lipid, and for (–3π/2, –π/2, π/2, 3π/2), (–π/2, π/2, 3π/2), (–5π/6, –π/6, π/2, 7π/6), and (–π/6, π/2, 7π/6). Note that the data were normalized such that the highest SNRfactor is 1. These plots show that the receiver BW dominates the effective noise performance, leading to a better SNRfactor at ±62.5 kHz, despite the fact that with an echo-shift combination of (–3π/2, –π/2, π/2, 3π/2) the NSA is less uniform than at (–5π/6, –π/6, π/2, 7π/6). It should be pointed out that with echo shifts (–5π/6, –π/6, π/2, 7π/6) more data are acquired in the phase-encoding direction, which improves spatial resolution. Also, with the echo-shift combination (–5π/6, –π/6, π/2, 7π/6), the timing differences between echoes (i.e., Δtn) is less than that for the echo-shift combination of (–3π/2, –π/2, π/2, 3π/2). Thus T2* effects for the echo-shift combination (–5π/6, –π/6, π/2, 7π/6) are less important, which may have an effect on the lipid-water separation.

thumbnail image

Figure 6. SNR factor of (a) water and (b) lipid for various water : lipid ratios and echo-shift combinations. The plots are normalized so that the highest SNRfactor = 1.

Download figure to PowerPoint

Lipid images of the lipid-water phantom obtained with the IDEAL-GRASE technique are shown in Fig. 7 for (a) (–3π/2, –π/2, π/2, 3π/2) and (b) (–5π/6, –π/6, π/2, 7π/6). Figure 7c shows the lipid image generated with the conventional chemical-shift saturation technique. It can be observed that images generated with the IDEAL-GRASE technique are superior in lipid-water separation over the conventional saturation technique. The lipid-water separation with echo shifts of (–5π/6, –π/6, π/2, 7π/6) is slightly better than (–3π/2, –π/2, π/2, 3π/2), which as explained above may be due to T2* effects.

thumbnail image

Figure 7. Lipid image of the lipid-water phantom obtained with IDEAL-GRASE with echo shifts (a) (–3π/2, –π/2, π/2, 3π/2) and (b) (–5π/6, –π/6, π/2, 7π/6). Data were acquired with a four-element phased-array coil with FOV = 20 × 20 cm2 and TR = 1 s. Other imaging parameters for (a) were BW = ±62.5 kHz, ETL = 8, matrix = 256 × 192 for each echo, and TEeff = 42 ms. Other imaging parameters for (b) were BW = ±125 kHz, ETL = 10, matrix = 256 × 240 for each echo, and TEeff = 35 ms. c: Lipid image obtained using the conventional chemical-shift suppression technique. The conventional FSE images were acquired with similar parameters to (a) except for ETL = 12.

Download figure to PowerPoint

Shown in Fig. 8 are the (a) water and (b) lipid images of the pelvis obtained with IDEAL-GRASE with echo shifts (–5π/6, –π/6, π/2, 7π/6). In addition to the water and lipid images, we can obtain the (c) in- and (d) out-of-phase images recombined from (a) and (b). When the calculated water and lipid images are recombined, the image distortion in pixel position and size (36) is not considered since it is negligible at 1.5T with BW = ±125 kHz (assuming 256 readout points) and moderate field inhomogeneity. Similar lipid-water separation can also be obtained from data with echo shift (–3π/2, –π/2, π/2, 3π/2) acquired at ±62.5 kHz (results not shown). The IDEAL-GRASE technique achieves better and more uniform lipid-water separation compared to the conventional saturation technique, which fails to suppress the lipid signal in the lower-left and upper-right regions of the water image, as indicated by the arrows in Fig. 8e.

thumbnail image

Figure 8. Pelvic (a) water, (b) lipid, and recombined (c) in- and (d) out-of-phase images obtained with IDEAL-GRASE. Data were acquired using a four-element phased-array coil with BW = ±125 kHz, echo shift = (–5π/6, –π/6, π/2, 7π/6), ETL = 10, acquisition matrix = 256 × 240 per echo, TR = 1 s, TEeff = 35 ms, NEX = 1, and FOV = 44 × 44 cm2. The scan time was 24 s. e: Water and (f) lipid images obtained with FSE using conventional chemical-shift saturation. The conventional FSE images were acquired with similar parameters except for ETL = 12. In the water image acquired with the conventional chemical-shift saturation technique, there is significant residual signal from lipid, as indicated by the arrows.

Download figure to PowerPoint

Figure 9 shows the (a) water and (b) lipid images of the knee obtained with the IDEAL-GRASE technique, the (c) water and (d) lipid images obtained with the IDEAL-FSE technique, and the (e) water image obtained with the conventional saturation method. Due to field inhomogeneities, the lipid suppression fails with the conventional saturation method, as indicated by the arrows in the bone and the back of the knee, where the suppression is incomplete and nonuniform. The IDEAL-GRASE technique successfully corrects for errors due to field inhomogeneities, achieving uniform and complete lipid-water decomposition. Lipid-water separation with IDEAL-GRASE is comparable to that achieved with the IDEAL-FSE technique.

thumbnail image

Figure 9. a: Water and (b) lipid images of the knee obtained with IDEAL-GRASE. (c) Water and (d) lipid images obtained with IDEAL-FSE. (e) Water and (f) lipid images obtained with FSE using conventional chemical-shift saturation. The arrows point to regions with incomplete or nonuniform lipid suppression. IDEAL-GRASE and IDEAL-FSE data were acquired with BW = ±125 kHz, echo shifts = (–5π/6, –π/6, π/2, 7π/6), ETL = 10, acquisition matrix = 256 × 240 for each echo, TR = 1 s, TEeff = 35 ms, NEX = 1, and FOV = 20 × 20 cm2, using a transmit/receive extremity coil. The FSE data with chemical-shift saturation were acquired with similar parameters except for ETL = 12. The total scan time for the acquisition of IDEAL-GRASE data was 24 s.

Download figure to PowerPoint

Water and lipid black-blood images of the heart obtained in a breath-hold by combining IDEAL-GRASE with a double-inversion preparation are shown in Fig. 10a and b, respectively. For comparison, the water and lipid images of the same slice generated from a double-inversion FSE (8) using the conventional chemical-shift saturation technique are shown in Fig. 10c and d, respectively. Both techniques produce similar lipid-water decomposition in the heart area. However, the conventional saturation technique yields incomplete separation in the liver (broken arrow) and in areas around the chest and back (solid arrows). This is not the case with the IDEAL-GRASE technique, which provides uniform lipid-water separation throughout the entire FOV. In this example we could not compare the results with those obtained by the IDEAL-FSE technique because with the latter, data cannot be acquired within a single breath-hold. One thing to note is that there is more residual signal from blood flow in the ventricular cavities of the IDEAL-GRASE image compared to the FSE image.

thumbnail image

Figure 10. Black-blood images of the heart of a normal volunteer acquired in a breath-hold using a GRASE or FSE technique combined with a double inversion preparation period for blood-flow suppression. a: Water and (b) lipid images obtained with the IDEAL-GRASE technique. (c) Water and (d) lipid images obtained with FSE using conventional chemical-shift saturation. The arrows point to regions with incomplete lipid-water separation. The IDEAL-GRASE data were acquired with BW = ±62.5 kHz, echo shifts = (–3π/2, –π/2, π/2, 3π/2), ETL = 8, acquisition matrix = 256 × 192 for each echo, TR = 1 RR, TEeff = 30 ms, NEX = 1, and FOV = 36 × 36 cm2, using a four-element phased-array coil. FSE data with chemical-shift saturation were acquired with similar parameters except for ETL = 12.

Download figure to PowerPoint

DISCUSSION

  1. Top of page
  2. Abstract
  3. TECHNIQUE
  4. MATERIALS AND METHODS
  5. RESULTS
  6. DISCUSSION
  7. CONCLUSIONS
  8. Acknowledgements
  9. REFERENCES

MRI based on FSE acquisitions is used routinely in the clinic, and the IDEAL algorithm was proposed to generate FSE data with robust lipid-water separation. We have shown that the combination of GRASE data acquisition with the IDEAL algorithm provides robust lipid-water separation in less time than with FSE. Furthermore, since the echoes for lipid-water separation are collected in an interleaved manner within each TR period, misregistration between data sets due to motion is minimized. Thus, the technique is suitable for imaging locations that suffer from motion artifacts or when scanning time is a constraint (e.g., as in breath-hold imaging).

The reduction in scan time compared to IDEAL-FSE depends on the parameters used. For example, for data acquired with BW = ±125 kHz and the echo-shift combination of (–5π/6, –π/6, π/2, 7π/6) for IDEAL-GRASE and (–π/6, π/2, 7π/6), which is the preferred echo-shift combination for IDEAL-FSE, the scanning time is reduced by a factor of 3. However, for data acquired with IDEAL-GRASE with BW = ±62 kHz, it is necessary to change the echo-shift combination to (–3π/2, –π/2, π/2, 3π/2). The latter echo-shift combination increases the time between adjacent 180° RF pulses by approximately 3 ms, forcing a longer duration of the echo train (and less slices per TR) or the acquisition of more TR periods (approximately 1.25 × TR periods) if one chooses to keep the duration of the echo train unchanged and a similar spatial resolution. In IDEAL-FSE the change of BW to ±62 kHz does not require a change in echo-shift combination (i.e., the echo-shift combination of –π/6, π/2, 7π/6 can still be used). Thus in this case, the time gain between IDEAL-GRASE and IDEAL-FSE is 2.4 rather than 3.

With the GRASE technique, it is necessary to flip the polarity of the readout gradients of adjacent echoes. The interaction of rapidly changing fields (such as those induced during fast switching of gradients) with conducting structures in the MRI scanner causes eddy currents (37). The conductive structures of the gradient coils themselves are a major cause of eddy currents. Other sources are the conductive structures on the RF coils. It was reported that the shielding that separates the transmit and receive circuits of transmit/receive RF coils is one source of eddy currents, and that the amount of eddy currents depends on the design of the RF shield (38). As shown in Fig. 2, the phase distortions are dependent on the RF coil, with the extremity coil yielding a greater phase distortion than the four-element phased-array coil. The extremity coil used in this work is not only a transmit/receive RF coil—it also has a multiple coil design (it is composed of a quad-birdcage, linear birdcage, and a saddle coil to cover the knee and ankle areas, and a solenoid to cover the foot and toes). This coil with multiple conductive elements seems to produce larger eddy currents. This explains the significant difference in lipid-water separation (prior to phase correction) between data acquired with the four-element phased-array coil (Fig. 3a and b) and the extremity coil (Fig. 3g and h). Eddy currents lead to field distortions that can have a nonlinear component leading to nonlinearity in the phase error, as indicated in Fig 2b for the case of larger FOVs. Other sources of phase distortions are differences in timing delays among the readout gradients used to acquire echoes E1E4. Errors due to timing delays lead to a shift in k-space that results in a linear phase term in image space. With the correction scheme proposed in this work, both linear and nonlinear phase errors are effectively removed, and lipid-water separation is complete regardless of the RF coil.

From the numerical simulations as well as the phantom results, we see that the acquisition of a fourth echo provides improved noise performance. Having extra echoes may also yield better lipid-water separation due to the extra information provided to the IDEAL algorithm. It is also possible to add more echoes, but this may introduce errors due to T2* effects. Therefore, in this work we chose to investigate up to four echoes with echo-shift combinations that do not increase the length of each SE period (and the total scan time) compared to the corresponding three-echo acquisition. The acquisition of more echoes may be possible if the iterative algorithm is modified to incorporate the effects of T2* (39).

One disadvantage of the method is that the choice of echo shifts is determined by the length of the readout window. This imposes constraints on the choice of echo-shift combinations, the number of readout points, or the receiver BW. This is not the case with IDEAL-FSE, where the echo-shift combination is independent of the readout window length. For example, the knee images presented in this work were acquired with 256 readout points, BW = ±62 kHz, and an echo-shift combination of (–3π/2, –π/2, π/2, 3π/2). Many clinical protocols use a higher number of readout points to increase spatial resolution. With IDEAL-GRASE it is possible to increase the number of readout points, and consequently the spatial resolution along the frequency-encoding direction, by adjusting the BW and the echo-shift combination. For example, data can be acquired with 512 readout points with the echo-shift combination of (–3π/2, –π/2, π/2, 3π/2) if the BW is doubled to ±125 kHz. For a different number of readout points, the BW and echo-shift combinations would have to be adjusted. Thus, a change in spatial resolution along the frequency-encoding direction is possible with IDEAL-GRASE, but the method is less flexible in terms of BW and echo-shift combination compared to IDEAL-FSE. Also, as noted above, GRASE may be more sensitive to artifacts from flowing spins compared to FSE. Further studies are needed to verify this observation and investigate possible solutions.

Integrated with the IDEAL algorithm, the region-growing method (35) leads to successful lipid-water separation in most applications by automatically choosing a suitable seed pixel. However, we found that in certain applications, such as cardiac imaging, misclassification occurs when an improper seed pixel is chosen. The failure of the region-growing method may be caused by disconnectivity or the sparsity of a high signal region in the heart image. Although one might be able to avoid this problem by manually choosing an appropriate seed pixel, it is necessary to find a more robust seed pixel searching strategy.

IDEAL-GRASE can be adapted for lipid-water separation at 3T, but the difference in timing between echoes, due to the increased chemical shift at 3T, needs to be accounted for. To achieve the same phase shift combinations used in this work at 1.5T, the length of the readout window needs to be adjusted. For example, to use the echo-shift combination of (–3π/2, –π/2, π/2, 3π/2) at 3T the BW has to be increased to ±125 kHz (this is still under the assumption of 256 readout points). To use a BW of ±64 kHz (while keeping the number of readout points unchanged), an alternative is to add a 2π increment to the echo shifts. Because a 2π increment is only 2.3 ms at 3T, the length of the acquisition window from E1 to E4 is only 1.1 ms longer than at 1.5T (for the same echo-shift combination). For other BWs and numbers of readout points, the optimal echo-shift combination needs to be found using an analysis similar to that illustrated in Fig. 6.

CONCLUSIONS

  1. Top of page
  2. Abstract
  3. TECHNIQUE
  4. MATERIALS AND METHODS
  5. RESULTS
  6. DISCUSSION
  7. CONCLUSIONS
  8. Acknowledgements
  9. REFERENCES

In this work we have demonstrated that robust and time-efficient water and lipid separation can be achieved by combining the GRASE method and the IDEAL algorithm. With the IDEAL-GRASE technique, four echoes can be acquired instead of three without increasing the scan time to achieve better noise performance and thus better lipid-water separation. With IDEAL-GRASE one complete data set can be acquired up to three times faster than with IDEAL-FSE without loss of spatial resolution; thus the proposed method represents a good choice when fast imaging is desired.

Acknowledgements

  1. Top of page
  2. Abstract
  3. TECHNIQUE
  4. MATERIALS AND METHODS
  5. RESULTS
  6. DISCUSSION
  7. CONCLUSIONS
  8. Acknowledgements
  9. REFERENCES

This work was supported by NIH grant CA099074 (M.I.A.) and Grant-in-Aid 0355490Z from the American Heart Association (M.I.A.). Student support was provided by the Arizona Hispanic Center of Excellence. The authors thank Dr. Angel Pineda for helpful discussions.

REFERENCES

  1. Top of page
  2. Abstract
  3. TECHNIQUE
  4. MATERIALS AND METHODS
  5. RESULTS
  6. DISCUSSION
  7. CONCLUSIONS
  8. Acknowledgements
  9. REFERENCES