Since fast spin echo (FSE), or rapid acquisition with relaxation enhancement (RARE) (1), is capable of generating high-quality T2-weighted images in a fraction of the time required by conventional spin echo (SE), it has largely become a sequence of choice for T2-weighted imaging. A prominently distinguishing feature of FSE, however, is its anomalously bright fat signal (2), a phenomenon that was attributed by Henkelman et al. (3) to the demodulation of the J-coupling and desensitization of diffusion through inhomogeneities due to the rapidly refocusing RF pulse trains.
Fat suppression is therefore a frequently desired imaging option in clinical T2-weighted imaging. Presently, the most commonly used method is chemical saturation (4), which, despite its many advantages, is known to be intrinsically susceptible to both the radiofrequency (RF) and the magnetic field inhomogeneity. An alternative technique that offers the potential of overcoming the problem is the Dixon technique (5). With this approach, the chemical shift difference between water and fat is encoded into images with different echo shifts. Field inhomogeneity effects appear as image phase errors, which in principle can be corrected for by a combination of multipoint acquisition and more elaborate image processing (6–10). While these techniques allow for more uniform water and fat separation in the presence of the field inhomogeneity, one obvious drawback is the requirement for multiple data acquisitions and therefore longer scan times.
Incorporating the Dixon approach for fat suppression into FSE presents a mutually beneficial combination. While the Dixon technique provides a potentially robust separation of the strong fat signal, FSE helps to alleviate the need for long data acquisition time in the multipoint Dixon technique. The feasibility of combining the two approaches was successfully demonstrated by Hardy et al. (11). In their implementation, however, echo shift as dictated by the Dixon technique is achieved by shifting the timing of the readout gradient and the data acquisition window in order to maintain the Carr-Purcell-Meiboom-Gill (CPMG) condition. As a result, interecho spacing is increased, which leads to a substantial loss in the slice coverage for a given sequence repetition time (TR), largely offsetting the gain of using FSE for reducing the scan time. Therefore, the technique is believed to be appropriate only for imaging small anatomic areas that do not require large slice coverage.
In this work, a new and more efficient method for FSE-based Dixon imaging was developed to achieve the echo shifts without necessitating an increase in echo spacing. In comparison to the previously-known method (11), increased slice coverage is possible within a given imaging time. Another important difference arises from the fact that the echoes collected in the modified method are generally asymmetric. It is shown that partial Fourier reconstruction methods can be used to preserve the phase information and to allow for generation of high-quality water-only and fat-only images, in both phantom and clinical patient imaging.
The echo shift in the original Dixon method (5) is achieved by time-shifting the 180° refocusing RF pulse. Alternatively, it can also be achieved by shifting the data acquisition window and the readout gradient while keeping the timing of the 180° pulse fixed (8, 11). The amount of the time shift in either case is directly proportional to the desired phase shift α between the water and fat magnetizations. However, for the same phase shift α, the time shift τ that is needed by the second method is twice that needed by the first method. At 1.5 Tesla, τ to achieve a phase angle difference α of 180° is approximately 2.3 ms.
Figure 1a displays the pulse sequence diagram of the FSE-based Dixon by Hardy et al. (11). In this FSE implementation, the interecho spacing needs to be increased by 2*τ, or twice the amount needed for shifting the data acquisition window and the readout gradient. Accordingly, a 4.6-ms increase in echo spacing is required to achieve a 180° phase shift between the water and fat signals at 1.5 Tesla field strength. Considering a typical 8-ms total readout time when using 256 readout points at 16 kHz receiver bandwidth (RBW), such an increase amounts to a significant lengthening of the sequence time and imparts increased image blurring due to the signal T2 modulation within the echo train. More importantly, the increase in echo spacing leads to a substantial protocol-dependent reduction, sometimes as much as 30–40%, in the slice coverage for a given imaging time.
The proposed diagram of the pulse sequence that was used in this work is illustrated in Fig. 1b. In comparison to the pulse sequence in Fig. 1a, a small gradient pulse (gxwl) with minimum duration is added prior to the readout gradient pulse (gxw). The net effect of the added gradient is to induce a spatially-linear spin phase shift along the readout direction, or equivalently a constant time shift in echo position. The area of the gradient pulse gxwl is set equal to the product of the amplitude of the readout gradient and the desired time shift in echo position. In order to preserve the CPMG condition, another gradient pulse of the same area, but of opposite polarity (gxwr), is added after the readout gradient. Since both gxwl and gxwr can be played out during the time set for slice crusher gradients and the phase-encode gradients, no timing change in RF, readout gradient, or data acquisition window is necessary. The spatially-linear phase shift along the readout direction induced by the added gradients is effectively restored by recentering the acquired data before image reconstruction (see next section). Consequently, neither an increase in echo spacing nor a loss in slice coverage is anticipated. It can be further noted that with the implementation in Fig. 1b, an arbitrary echo shift can be effected simply by changing the areas of the gradient pulses gxwl and gxwr without changing the timing of any other pulses. In fact, with a fixed pulse width for gxwl and gxwr, echo shift is directly proportional to, and conveniently controlled by the amplitude of the gradient trapezoids.
Although the proposed pulse sequence is compatible with various multipoint-Dixon acquisition techniques, the three-point asymmetric sampling scheme with 0°, 90°, and 180° phase shifts (10) was implemented for this work. Apart from providing increased timing flexibility, the asymmetric sampling has been shown to offer increased processing reliability (10). In order to restore the spatially-linear phase shift induced by the added gradients, echoes with non-zero phase shifts are first recentered by an amount equal to the echo shifts induced by the gradient gxwl before the image reconstruction. As a result, the raw data for these echoes are asymmetric, with a missing portion that is proportional to the phase shifts between the desired water and fat magnetizations.
The asymmetric data structure is similar to that of a fractional echo, for which many image reconstruction algorithms have been proposed (12). However, most of the algorithms are only valid under the assumption that the object being imaged is real, and therefore are not directly applicable to the Dixon technique, whereby relative phases are intentionally introduced between the water and fat magnetizations. In this work, we used the homodyne reconstruction algorithm (13), but care was taken to preserve the image phase information by first performing the following Fourier transform:
where L(kx, ky) and H(kx, ky) represent the central symmetric and outer asymmetric portions of the acquired data, respectively; u(kx) is a unit step function that is used to effectively double the weight of the asymmetric portion of the data in the Fourier transform; ϕ(x,y) is the spatially varying phase error that arises from factors such as the gradient timing and RF imperfections; and m(x,y) and h(x,y) are the desired and the high-resolution components of the image, respectively. Equation  is valid if ϕ(x, y) can depict adequately the phase of the images reconstructed using either the full or only the central portions of the k-space data (13). It should be noted that in addition to the desired image m(x,y)eiϕ(x,y), there is generally a blurring term that is equal to the convolution of h(x,y)eiϕ(x,y) with a kernel 1/iπx. Obviously, the contribution of the blurring component is related to the amount of the missing portion of the acquired data. When the phase shift is 0, the acquired echo becomes symmetric and Eq.  is reduced to a regular Fourier transform.
Since the phase error term eiϕ(x,y) is usually slow-varying spatially, it can be adequately determined from low-resolution images without compromising the final image resolution. As a matter of fact, it has been demonstrated that faster and more reliable determination of the phase error terms can be achieved in Dixon processing with low-resolution images because they have reduced matrix size and increased signal-to-noise ratio (SNR) (14). In the present work, three sets of low-resolution images (corresponding to the acquisitions with 0°, 90°, and 180° phase shifts) were reconstructed from the central symmetric portion (either 128 × 128 or 64 × 64) of the acquired multipoint Dixon data. A region-growing algorithm without direct phase unwrapping, similar to that in Ref. 15, was employed to determine the phase error term eiϕ(x,y) from these three low-resolution images corresponding to the different echo shifts.
Assuming eiϕ(x,y) does not vary significantly over the scale of 1/iπx, it can then be used to demodulate the image in Eq. :
Note that in the regular homodyne reconstruction, both m(x,y) and h(x,y) are assumed to be real, and the blurring component (the second term in Eq. ) can be discarded simply by taking the real part of Eq. . In Dixon imaging, it is recognized that this can also be performed when the phase angle difference is 0° or 180° because the water and fat magnetizations are then along the same axis. For phase angles other than 0° or 180°, the object is generally not real and there is usually an intermix of the blurring component and the desired image in the real and imaginary channels. Under such circumstances, the blurring component can in principle be estimated using an approach proposed in Ref. 16, or some more elaborate iterative methods (12). However, in our application, since only 36 out of a typical 256 data points were not acquired at 16 kHz RBW, simple zero-filling of the data and direct Fourier transform were used to obtain the image corresponding the 90° phase shift.
A product FSE pulse sequence by GE Medical Systems (Waukesha, WI) was modified to make two separate pulse sequences, as illustrated in Fig. 1a and b. Both pulse sequences were compiled under the 83M5 software release for GE 1.5 Tesla Signa Lx systems. A phantom consisting of a bottle of Hytop vegetable oil (Federated Group, Inc., Arlington Heights, IL) and a bottle of distilled water solution with copper sulfate pentahydrate and sodium chloride were imaged using both pulse sequences to compare their performance and time efficiency. For in vivo demonstration, both pulse sequences in Fig. 1 were used to image the spine of a clinical patient under the same imaging conditions. In addition, the pulse sequence in Fig. 1b was used to image the head/neck and orbital regions of two other clinical patients, for which the large slice coverage required makes the pulse sequence in Fig. 1a less practical. In all situations, comparison was made on the same patients at the same scan location to the product FSE sequence with chemical saturation for fat suppression, which is part of the current clinical protocol at our institution. All human studies were conducted with approval by the University of Texas M.D. Anderson Cancer Center Institutional Review Committee, and informed consents were obtained from all subjects.
Reconstruction of the separate water-only and fat-only images was performed using the raw data saved from the scanners (all 1.5 Tesla) and on an IBM PC under the Microsoft Windows 2000 platform. The reconstruction algorithm was implemented in Matlab (Math Works, Natick, MA). After reconstruction, images were transferred via the hospital network back to the scanners and installed into the image database for archiving and filming. Other than requiring some input on patient-dependent image header information, the entire reconstruction process was automatic and did not need any user intervention.
The T1 and T2 of the water solution in the phantom were measured to be 460 ms and 380 ms, respectively. A single-voxel spectroscopy pulse sequence with a TR of 4000 ms, a TE of 85 ms, and center-frequency on fat was used to obtain a spectrum of the vegetable oil (Fig. 2), which shows the presence of the olefinic protons, in addition to the other two major peaks corresponding to the methylene and the terminal methyl protons. While the methylene and the terminal methyl proton resonance peaks are close to each other, the olefinic proton peak is approximately 260 Hz away from the CH2 and CH3 peaks, and thus lies very close to the water resonance.
An exemplary phantom image using the pulse sequence in Fig. 1b of the first Dixon acquisition (when the phase shift is 0), and the processed water-only and fat-only images of the same slice are displayed in Fig. 3a–c, respectively. The following imaging parameters were used: TR = 3400 ms, TE = 85 ms, echo train length (ETL) = 12, RBW = 16 kHz, acquisition matrix = 256 × 256. A total of 21 slices were collected in 3:51 min. Using the pulse sequence in Fig. 1a, only 15 slices were possible when all the imaging parameters were kept the same. Figure 3d–f shows the corresponding image of the first Dixon acquisition and the processed water-only and fat-only images of one of the 15 slices corresponding to the same spatial location as in Fig. 3a–c, but using the pulse sequence in Fig. 1a. In order to image 21 slices, TR would have to be increased to 4600 ms, and the total imaging time would have become 5:04 min for the pulse sequence in Fig. 1a.
The SNR, defined as the mean signal of a region-of-interest (ROI) over the mean amplitude of background, for the water region in Fig. 3a is 87. For the same location in Fig. 3b, SNR is 205. These SNR values are similar to those measured from the same location in Fig. 3d and e, which are 79 and 222, respectively. The increase in SNR of the image in Fig. 3b over the image in Fig. 3a, or of the image in Fig. 3e over the image in Fig. 3d, can be attributed to a combined gain due to the phase-sensitive reconstruction (17) and the multiple acquisition of the Dixon technique (approximately an equivalent of three excitations) (10). The residual fat signal, defined as the signal intensity of a fat region in the water-only image over the signal intensity of the same region in the fat-only image, is found to be 14.8% and 15.0% for the pulse sequences in Fig. 1a and b, respectively. The presence of the finite amount of the fat signal in the water-only images in both Fig. 3b and e can be attributed to contribution from the olefinic protons in the fatty acid, and agrees reasonably well with an estimation of 20% by Kuroda et al. (18) based on spectrum fitting. In contrast, the residual water signal, defined as the signal intensity of a water region in the fat-only image over the signal intensity of the same region in the water-only image, is only 0.5% for both pulse sequences in Fig. 1a and b.
Figure 4a shows a T2-weighted FSE image with conventional chemical saturation of a patient with metastatic breast cancer. The following imaging parameters were used: TR/TE/ETL = 3100 ms/102 ms/16, acquisition matrix = 384 × 256, RBW = 16 kHz, FOV = 32 cm, 4 NEX or signal averages, superior and inferior spatial saturation. A total of 10 slices were acquired in 3:25 min. The uneven fat saturation observed superiorly in the cervico-thoracic junction is not that uncommon at our institution when cervical and/or thoracic spines are imaged with chemical saturation. Figure 4b and c show the water-only images of the same slice using the three-point Dixon techniques, as shown in Fig. 1a and b, respectively. When all imaging parameters were kept the same except for a TR of 2950 ms and 2 NEX (for an equivalent of six signal averages), 10 slices were collected in 4:49 min with the pulse sequence in Fig. 1b. In comparison, only seven slices could be acquired with the pulse sequence in Fig. 1a within the same total time. Both images in Fig. 4b and c show more uniform fat suppression over the entire FOV than the image in Fig. 4a, resulting particularly in better depiction of the cervico-thoracic junction and the posterior elements of the vertebral bodies. The overall quality of the fat suppression is similar between the two Dixon acquisition techniques. However, careful examination reveals that the image in Fig. 4c appears visually crisper, such as in the spinal cord. Such a difference could conceivably be a consequence of the reduced interecho spacing for the pulse sequence in Fig. 1b as compared to that in Fig. 1a.
Figure 5a shows a T2-weighted FSE image with conventional chemical saturation of the sinus region of a second patient. The image is one of the 18 slices required to cover the head and neck region of clinical interest, and was obtained in 4:42 min with the following imaging parameters: TR/TE/ETL = 3625/90/12, acquisition matrix = 256 × 224, RBW = 16 kHz, FOV = 16 cm, and 4 NEX. Figure 5b shows the water-only image of the same slice as in Fig. 5a using the three-point Dixon technique as shown in Fig. 1b. With all the imaging parameters kept the same except for a TR of 3150 ms and 1 NEX (for an equivalent of three signal averages), 18 slices were collected in 3:16 min. By contrast, if the pulse sequence in Fig. 1a were used to achieve the same slice coverage, the scan time would have become 4:22 min. Comparison of the image in Fig. 5b to that in Fig. 5a again shows an overall improvement in fat suppression, particularly in the suppression of the premaxillary subcutaneous fat. Furthermore, less motion artifact and clearer anatomic delineation are observed, possibly as a result of the shorter imaging time employed.
Figure 6 shows a comparison of two head images of another patient through the orbital regions from the conventional chemical saturation (Fig. 6a) and from the modified Dixon technique (Fig. 6b). For the chemical saturation technique, 23 slices were acquired in 5:08 min with TR = 4675 ms, TE = 90 ms, ETL = 12, RBW = 16 kHz, three signal averages, and acquisition matrix = 256 × 256. The performance in fat suppression is actually comparable in both Fig. 6a and b. However, a shorter TR (4150 ms) can be used for the modified Dixon technique because no chemical saturation pulses are needed. As a result, the same 23 slices were acquired in 4:34 min with all other imaging parameters kept the same.
DISCUSSION AND CONCLUSIONS
A practical limitation to clinical use of the multipoint Dixon technique is its long acquisition time, particularly for T2-weighted imaging. Combining FSE with the Dixon technique is effective in reducing the total scan time, just like FSE is used to reduce the scan time of the conventional SE. However, due to the increased echo spacing, the previously-reported FSE-based Dixon technique exerts a severe loss in slice coverage, which limits its clinical application. The method proposed herein circumvents these drawbacks and can make the FSE-based Dixon technique as efficient in slice coverage as regular FSE. In fact, for the same TR, the new FSE-based Dixon approach usually allows for a few more slices than the regular FSE because chemical saturation pulses are eliminated. Practically, more time is saved with the FSE-based Dixon imaging because there is no need for a manual prescan and/or even repeat scans, as is sometimes required when using the chemical saturation. Considering the preservation of SNR per unit time of the multipoint Dixon acquisition, the present implementation can be used efficiently and effectively to image many different anatomic areas.
One inconvenience with the three-point Dixon technique that was occasionally encountered is that the combination with an imaging option to suppress the image wraparound artifact in the phase-encode direction is not possible without increasing the actual number of averages to six. In regular imaging, artifact suppression with this imaging option is achieved through doubling the FOV and the number of phase-encode steps. The total scan time is maintained by using an actual number of averages that is half of what is prescribed. Since reconstruction of Dixon images is not compatible with half k-space data along the phase-encode direction, Dixon imaging with such an artifact suppression option forces a doubling of the scan time as compared to the actual minimum required scan time. This was the underlying consideration for the disparity in the scan time selected for Fig. 5a and b. Using an effective number of an average of six would increase the total scan time to over 6 min, potentially exacerbating the motion degradation that is visible in Fig. 5a. One way to alleviate the problem is to investigate the feasibility of a Dixon image reconstruction algorithm that is compatible with data that fills only a fraction of the k-space along the phase-encode direction. With the proposed Dixon acquisition strategy, however, the image reconstruction algorithm would have to deal with some data that are fractional along both the frequency and phase-encode directions. A different approach to minimize the scan time and to avoid the phase aliasing is to exploit the inherent encoding effect in the sensitivity maps of a phased-array coil, as in some of the recently-proposed parallel imaging methods (19, 20). In this case, coil sensitivity and gradient would be used to encode spatial information while the echo shift would encode the chemical shift between water and fat.
In comparison to the regular FSE, the pulse sequence in Fig. 1b may introduce additional diffusional weighting due to the added gradient pair (gxwl and gxwr), and may affect gradient moment nulling along the readout direction—an option that is sometimes used for reducing flow and motion artifacts. While it is possible to design a gradient for gxwl and gxwr that satisfies the gradient moment nulling condition, the more complicated waveform will naturally lead to longer gradient duration and thus offset the benefit of the pulse sequence for reduced echo spacing. Nonetheless, since the area of the added gradient lobe is relatively small (<15% of the readout gradient), the flow and diffusional effects they introduce are expected to be limited. Compared to the sequence in Fig. 1a, the reduced interecho spacing should help to compensate to a certain extent any added flow and diffusional spin dephasing.
Variations in the implementation shown in Fig. 1b are certainly possible to effect the similar echo shifts. One possibility is to use a negative gradient lobe before the readout gradient and a positive gradient lobe after the readout gradient. In this case, the echo center will be shifted to occur after the Hahn echo. Although they are seemingly equivalent, echoes collected using this approach would have a slightly larger T2 decay and also contain potential contamination from the free induction decay (FID) signals due to the imperfect 180° refocusing pulses. It is important to note that regardless of how echoes are shifted, the time shift is determined by the phase shift desired. When this is held constant, the total readout time, which is determined by the readout bandwidth and the number of readout points, determines the relative portion of the missing echoes. Apparently the method is not expected to work well if the missing portion is close to half of the total echo. In that case, echo spacing would have to be increased to accommodate the echo shift, as is done in Fig. 1a.
In conclusion, we have developed a new and more efficient FSE-based Dixon acquisition technique that does not require increased echo spacing. Combined with phase-sensitive partial Fourier image reconstruction and Dixon processing, the technique was capable of generating high-quality water-only and fat-only images in both phantom and in vivo patient imaging. With increased time efficiency, the technique can be used to expand the clinical application of the multipoint Dixon technique, and will provide an attractive alternative to the more widely-used chemical saturation technique when the performance of the latter approach is not satisfactory.
Krista McAlee is acknowledged for her assistance and excellent support in acquiring patient images. Helpful discussions with Dr. X.J. Zhou, and assistance from Dr. E.F. Jackson in obtaining the fat spectrum are also appreciated.