Spiral water–fat imaging with integrated off-resonance correction on a clinical scanner




To integrate water-fat–resolved spiral gradient-echo imaging with off-resonance correction into a clinical MR scanner and to evaluate its basic feasibility and performance.

Materials and Methods

Three-point chemical shift imaging was implemented with forward and strongly T2*-weighted reverse spiral sampling and with off-resonance correction after water–fat separation. It was applied in a volunteer study on single breathhold abdominal imaging, which included a brief comparison with Cartesian sampling.


Water-fat–resolved, off-resonance–corrected forward and reverse three-dimensional interleaved spiral imaging was found to be feasible on a clinical MR scanner with only minor changes to the existing data acquisition and reconstruction, and to provide good image quality. Three-point chemical shift encoded data thus support both, water–fat separation and off-resonance correction with high accuracy.


The combination of chemical shift encoding and appropriate postprocessing could pave the way for water-fat–resolved spiral imaging in clinical applications. J. Magn. Reson. Imaging 2010;32:1262–1267. © 2010 Wiley-Liss, Inc.

THE DIFFERENTIATION BETWEEN water and fat is often essential to improve image contrast in MRI. In abdominal applications, it is of special importance, because the fat signal contribution can obscure relevant pathologies (1). However, the fat signal itself also provides important diagnostic information, for example, perilesional fat sparing can help to detect tumors (2), and the amount and the distribution of fat is of direct interest in obesity-related diseases (3). Therefore, several methods have been proposed to distinguish between the two species. Chemical shift selective approaches (4), measuring either water or fat, have been developed, but often suffer from main field inhomogeneities or local tissue susceptibility variations. By contrast, chemical shift encoding approaches, including three-point Dixon methods (5, 6), seem to be more robust against spatial variations of the main field B0 and allow the simultaneous measurement of both species.

Water–fat resolved clinical imaging is currently dominated by Cartesian sampling schemes (5–7). While alternative, non-Cartesian approaches, such as spiral sampling schemes (8), have much to offer in terms of sampling efficiency, choice of echo times (8, 9) and tolerance to motion (8), they show a high sensitivity to off-resonance (10). Image blurring artifacts, either resulting from spatial variations of B0 and/or chemical shift composition, degrade image quality seriously. Recently, Moriguchi et al (11) and Brodsky et al (12) showed that if these two sources of off-resonance are separated, image deblurring is possible. Three-point Dixon approaches (5), including the recently developed variant of IDEAL (6, 7), facilitate such a separation, both by distinguishing water and fat signal contributions and by providing a map of ΔB0, the deviation from the nominal main field strength B0. Therefore, this approach could be an interesting element to improve image quality with spiral sampling.

In this work, this concept was implemented and evaluated on a clinical MR system. The existing spiral acquisition and reconstruction software was slightly modified and off-resonance correction was integrated into the reconstruction software. These changes facilitate water-fat–resolved, off-resonance–corrected two-dimensional (2D) and 3D forward and reverse spiral imaging in a clinical environment. This allows analyzing the performance of spiral imaging in a variety of clinical applications, without compromising image quality or delaying reconstruction unacceptably.

The basic concept and the changes made to the acquisition and reconstruction software will briefly be described. Basic feasibility will be demonstrated on the example of 3D single breathhold abdominal imaging using forward and reverse spiral sampling. Furthermore, a brief comparison to a selected Cartesian water-fat–resolved scanning approach will be made.


Water–Fat Separation and Off-Resonance Correction

To ease sequence design and to simplify image reconstruction/correction, a fixed k-space trajectory is assumed for the chemical shift-encoded spiral data acquisition (see Fig. 1a,b). Different images corresponding to different echo times TEn (n ∈ [1, N]) are reconstructed using conventional gridding. The image signal Sn(r) for a given pixel at location r can consequently be written as a superposition of the different signal contributions ρj of the different chemical species denoted by the index j

equation image(1)

The spatial signal distributions ρj of the different species are convolved with the spatially variant point-spread-function (PSF), reflecting the blurring induced by spiral sampling in presence of off-resonance (main field inhomogeneity ΔB0(r), chemical shift ωj). The mathematical structure of the signal model given in Eq. [1] is identical to the one derived for conventional Cartesian signal sampling (5, 11, 12). Consequently, existing chemical shift decomposition approaches working in the spatial domain can be used to estimate ρj ′(r). In this work, an IDEAL-type approach (6) is chosen, which allows, based on three differently chemical shift-encoded echoes, separating the spectral signal components of water and fat, additionally delivering a spatially resolved off-resonance map (ΔB0).

According to Eq. [1], the resulting separated water and fat signals are still blurred by local PSF contributions, which can be removed in a subsequent step using a conjugate phase reconstruction (CPR) (10, 13). For the water signal, which is assumed to be on-resonance, CPR using the ΔB0 map is sufficient, whereas the fat signal has to be de-blurred additionally for the chemical shift offset ωj. This procedure, called DIXON-CPR in the following, is schematically shown in Figure 1d. More sophisticated approaches are conceivable, but due to its simplicity the present work focuses on this two-stage approach, which has the potential to be easily applicable in a clinical setup.

Figure 1.

Chemical shift encoding and water/fat separation. a,b: Schemes of three-point Dixon spiral sampling with the spiral trajectory kept fixed, but shifted in time to encode the off-resonance. a: Forward spiral. b: Reverse spiral. c: Scheme of the multi-coil reconstruction. After reconstructing the images for each coil individually, images are combined by a signal phase preserving Roemer/SENSE reconstruction, yielding one complex spiral image per TE for efficient postprocessing. d: Scheme of the spiral water–fat separation and correction algorithm (DIXON-CPR).

So far, single coil reception has been assumed. In case of multi-coil signal reception, the water–fat separation has to be performed for each individual channel (6), followed by an appropriate CPR. This is a huge computational burden. To avoid it, the individual multi-coil images are merged in the present implementation into a single image using a Roemer (14) or, in case of accelerated scanning, a SENSE reconstruction (15) before performing further processing (c.f., Fig. 1c). This image combination is optimal in terms of SNR (14, 15) and preserves the signal phase, i.e., the chemical shift encoding information imprinted in the data is not lost. Thus, the DIXON-CPR step is only applied once at the end of the reconstruction chain to the fused data, which reduces the computational effort substantially (Fig. 1c,d).

Experimental Methods

To test this approach, abdominal imaging experiments were performed on a 1.5 Tesla (T) clinical scanner (Achieva, Philips Healthcare, Best, NL) using a 32-element cardiac coil (In-Vivo Corporation, Gainesville, FL). The reconstruction software of the scanner was slightly modified and CPR, according to the approach described in Eggers et al (13), was added after the existing water–fat separation module (c.f., Fig. 1c,d). Furthermore, spiral data acquisition was extended to allow reverse (9) spiral sampling. Nine healthy volunteers (all male, 33–50 years) were recruited for this study, and informed consent was obtained according to the rules of the institution. The 3D interleaved forward and reverse spiral gradient echo imaging was performed using the stack-of-spirals approach (16), using a variable angular speed trajectory. Stack-of-spirals sampling was performed in transversal orientation using a 30% oversampling in the slab phase encoding direction, compensating for radiofrequency (RF) pulse imperfections. To increase volume coverage in a given time and to compensate for the scan time increase due to chemical shift encoding, conventional SENSE (15) acceleration was applied in the stack, that means in the superior/inferior (S/I), direction. The sequence parameters for the different scans performed are summarized in Table 1.

Table 1. Parameters of the Different 3D Single Breathhold Scans (of 24 s duration)
 TypeMatrixFOV (mm3)AQ/TR (ms)interl.TR (ms)αSENSE TEi (ms)
  1. Scans (I–IV) are spiral scans with SENSE applied only in S/I direction. Scan (V) is a multi-gradient echo Cartesian scan, which is also slightly accelerated in the anterior-posterior direction. Note that spiral scans sample a net circular, not square, region of kxky space. A similar reduction was also applied in the Cartesian sampling for the 2D phase encoding k-space. From the rectangular phase encoding sub-k-space, only 78.5% was sampled. The table lists scanning type, matrix, FOV, acquisition window per TR, number of interleaves (only for spirals), repetition time, tip angle, SENSE reduction factor, and echo times. The total signal sampling time per scan, useful as a measure for the noise power received, was comparable for all sequences, ranging between 13 and 15 s.

(I)forward256 × 256 × 46400 × 400 × 23015132315°2.41.0/2.5/4.0
(II)reverse256 × 256 × 37400 × 400 × 18515172315°2.417.0/18.5/20.0
(III)forward240 × 240 × 54400 × 400 × 21615132315°2.71.0/2.5/4.0
(IV)forward240 × 240 × 54400 × 400 × 21610171815°2.71.0/2.5/4.0
(V)mGE240 × 192 × 54400 × 320 × 2163 × 1.36.610°2.7 × 1.11.6/3.3/5.0

During the 3D acquisition, the sequence was repeated, for each spiral interleaf three-fold, shifting the corresponding spiral read-out by a temporal increment of 1.5 ms to address different TEs, keeping TR fixed. Chemical shift encoding is thus performed in the innermost loop of the data acquisition, minimizing motion- and flow-related data inconsistencies.

The protocol performed in each volunteer comprised five different scans (see Table 1), each performed in a single breathhold of 24 s. The 3D forward and reverse spiral imaging using a moderate SENSE acceleration factor was performed (scan I, II) to show basic feasibility. For comparison, both protocols used the same voxel size, but for the reverse spiral, slightly smaller S/I coverage was used to relax the gradient demands (scan II). Two additional 3D forward spiral scans with slightly modified spatial resolution were added (scan III and IV). Both had the same spatial resolution and volume coverage, but differed in the spiral sampling window, allowing studying the impact of residual blurring, that cannot be compensated by the CPR.

To compare potentially remaining spiral off-resonance artifacts, an efficient Cartesian scheme was added (scan V). In this multi-echo sequence, three gradient echoes were sampled after each RF excitation using an alternating read-out gradient, allowing high sampling efficiency (17). An EPI-type phase correction was applied to remove eddy-current induced odd/even echo phase inconsistencies. The high pixel bandwidth used in this experiment (0.77 kHz) kept the odd/even related fat signal shift within a fraction of a pixel, which was neglected, i.e., no corrections were applied.

Spiral image reconstruction, SENSE unfolding and the following DIXON-CPR were performed on the scanner's reconstruction hardware with slightly modified software. DIXON processing was performed in 3D, using a 5 × 5 × 5 pixel field map smoothing kernel. As a consequence the spatial resolution of the estimated ΔB0 map is slightly reduced, restricting the ability of the following CPR algorithm to remove blurring. However, CPR is based on the assumption of a spatially smooth field map anyway (10), justifying the choice of the kernel size as a good compromise. The corresponding results were visually inspected and judged by two experienced readers using a score system ranging between 1 and 5 (very good–not acceptable). Water–fat separation performance, image quality and residual blurring were judged. To evaluate residual spiral blurring, a comparison with the corresponding results obtained with the Cartesian scheme was particularly helpful. Therefore, also the Cartesian results were analyzed with respect to water-fat–separation performance and image quality.


All volunteer scans yielded good and reproducible image quality. The image quality for the water/fat separated and off-resonance deblurred forward and reverse spiral scans was found to be good (2.1 ± 0.4) on average. Figure 2 shows representative forward spiral data, using a sampling window of 15 ms. Blurring present in the single echo images sampled at different TEs (Fig. 2a–c) was successfully removed in the final, separated water and fat images (Fig. 2g,h). To illustrate the procedure, intermediate results are shown in Figure 2d–f. This field map-based de-blurring procedure improved image quality throughout the entire volume, which is demonstrated in Figures 3 and 4. Among all volunteers studied, the global 3D ΔB0-induced frequency shift was found to be 0.34 kHz ± 0.05 kHz (standard deviation), showing a strong gradient in FH direction. The maximum in-plane variation among all transversal slices in the 3D volume, which is most relevant for CPR, was lower (0.22 kHz ± 0.05 kHz) and comparable to the chemical shift of fat. For the compensation of such an off-resonance, approximately 8 bins were necessary in the CPR correction (10, 13). Heavily T2* weighted reverse spiral data are shown in comparison to corresponding forward spiral data in Figure 3. Slight artifacts are visible in the reverse spiral data, especially in the fat images as small rims or slight ghosting near the chest wall (c.f., Fig. 3b,d), which can be attributed to minor eddy currents.

Figure 2.

Single breathhold water– fat resolved abdominal 3D forward spiral imaging. a–c: Selected transversal slice from a 3D data set measured at different TEs (1.0/2.5/4.0 ms). Off-resonance artifacts are visible. d–f: Corresponding water–fat images and ΔB0 map after DIXON separation (c.f., flow chart in Fig. 1c,d). g,h: Water and fat images after CPR.

Figure 3.

Single breathhold water– fat resolved abdominal 3D forward/reverse spiral imaging. a: Selected water image after DIXON-CPR of a forward spiral acquisition (TEeff = 2.5 ms). b: Corresponding water image of a heavily T2*-weighted reverse spiral acquisition (TEeff = 18.5 ms). c,d: Corresponding fat images. e–h: Corresponding coronal reformats, to illustrate the volume coverage.

Figure 4.

Single breathhold water – fat resolved abdominal 3D forward spiral and multi-echo Cartesian imaging. a–c: Selected images forthe different sampling schemes: spiral images after DIXON-CPR using a sampling window of 15 ms (left column) and 10 ms (middle column), Cartesian images after DIXON only (right column). a–c: Water images. The arrows highlight the different appearance of a susceptibility-induced artifact. d–f: Corresponding separated fat images for the same slice. g–l: Corresponding coronal reformats of the water and fat images.

Figure 4 shows representative results for spiral scans using different sampling windows in comparison to a Cartesian scan. In general, no serious differences in image quality were found, but Cartesian images were found to be on average (1.7 ± 0.4) slightly better than spiral ones (given above). This change was found to be almost statistically significant (P value: 0.06). The main reason for this difference is the residual spiral blurring, which could not completely be removed. Especially in the intestinal area some small blurring remained, which is reflected in the de-blurring performance, which was scored to be good (1.9 ± 0.2), but not excellent. Also after CPR in the vicinity of strong susceptibility gradients, present at tissue-air interfaces, hot-spot artifacts or small residual image blur are visible in the spiral data (c.f., Fig. 4 a–c), unlike in the Cartesian data. However, the water-fat–separation quality was significantly better for the spiral scans (1.1 ± 0.2) compared with the multi-echo Cartesian scans (1.5 ± 0.1), which showed some separation problems at the edges of the imaging volume due to residual eddy current problems.

The changes made to the reconstruction software slightly prolonged image reconstruction. For the measured 3D spiral data sets, water–fat separation took approximately 12 s, whereas CPR for water and fat images took on average 12 s in total, at the given range of field inhomogeneities. Overall, water-fat–separated and off-resonance–corrected spiral images were accessible for the user approximately 35 s after completion of the 32 channel 3D data acquisition.


The experiments confirmed that high resolution, off-resonance-corrected, chemical shift-encoded water-fat–resolved spiral imaging is feasible on a clinical MR scanner, requiring only minor modifications to the scanner's data acquisition and reconstruction software. Such a modified MR scanner can serve as a research platform for studying spiral imaging applications clinically, without compromising image quality. The appealing aspect of multi-point Dixon approaches is the estimated field map that facilitates deblurring, resulting in remarkable image quality improvements. This allows to perform high quality forward and heavily T2*-weighted reverse spiral imaging with long data acquisition windows for high sampling efficiency. In this way, the DIXON-CPR approach facilitates water-fat–resolved abdominal imaging, as investigated here, and potentially other spiral imaging applications.

The modified reconstruction software for multi-coil data allows ROEMER/SENSE image combination before applying the DIXON-CPR water/fat separation and image correction as a single and efficient postprocessing step. This postprocessing step can further be extended, if corrections for k-space trajectory imperfections become necessary.

In this study, a fixed field map filtering kernel was used in the water/fat separation. This might not be optimal, and further studies on the interplay of the applied B0 map smoothing and CPR accuracy are needed (11). Potentially, an iterative procedure could be interesting, helping to refine the ΔB0 estimation after having water and fat sufficiently separated (18). This might reduce the sensitivity of the spiral to local susceptibility gradients as shown in Figure 4a–c. The local gradients are not sufficiently captured by the field map, thus CPR was not able to remove the corresponding artifacts. The effect of these artifacts scales with the spiral sampling window duration, and an appropriate compromise has consequently to be found (10).

Furthermore, the multi-line nature of the fat signal should be considered (12, 19) in the separation and CPR to suppress the corresponding chemical shift-induced image blurring more fully. Yu et al (19) proposed to modify the signal model in the IDEAL algorithm, fitting also a signal decay time constant (T2*) for water and fat. The resulting T2* map could potentially be used in a correction, similar to CPR, reducing T2*-induced pixel blurring. However, given the limited maximum chemical shift encoding time used, compared with the spiral acquisition windows used, the accuracy of this T2* estimation will be rather limited, making such a correction questionable.

For the abdominal application studied here, the spiral also shows some weaknesses, however. The transversal body cross-section often favors rectangular FOV sampling, which is much easier to exploit with the Cartesian scheme. Furthermore, Cartesian schemes show also much more simplicity regarding parallel imaging. The reduction factors used in this study were dictated by the spiral scans to allow a comparison. Spiral subsampling would also be desirable in the in-plane directions, but due to its numerical complexity, this is currently out of reach for clinical applications.

The current study shows that DIXON-CPR–based spiral imaging results in good image quality. However, the question whether the spiral or a Cartesian sequence should be used depends mainly on the application. Further improvements in scan efficiency to the proposed sampling scheme are conceivable using two-point Dixon (20) and/or compressed sensing (21) approaches. In this way, the volume coverage and the spatial resolution can be increased or the total scan time, especially the not patient friendly breathhold duration, can be reduced.

In conclusion by small modifications to the data acquisition and reconstruction software of clinical MR scanners, water-fat–resolved, off-resonance–corrected forward and reverse 3D spiral imaging becomes feasible at high image quality. Such a platform is attractive to further study and evaluate the benefits of spiral sampling in a variety of potential applications. A single scan delivers all information necessary to support water–fat separation and facilitates image self-deblurring. Such an approach can pave the way for use of efficient spiral imaging in future clinical applications.