|Home | About | Journals | Submit | Contact Us | Français|
Balanced steady-state free precession (bSSFP) MRI is a rapid and SNR-efficient imaging method, but suffers from characteristic bands of signal loss in regions of large field inhomogeneity. Several methods have been developed to reduce the severity of these banding artifacts, typically involving the acquisition of multiple bSSFP data sets (and the accompanying increase in scan time). Fat suppression with bSSFP is also challenging; most existing methods require an additional increase in scan time, and some are incompatible with bSSFP band-reduction techniques.
This work was motivated by the need for both robust fat suppression and band reduction in the presence of field inhomogeneity when using bSSFP for flow-independent peripheral angiography. The large flip angles used in this application to improve vessel conspicuity and contrast lead to SAR considerations, longer TR, and increased severity of banding artifacts. In this work, a novel method that simultaneously suppresses fat and reduces bSSFP banding artifact with the acquisition of only two phase-cycled bSSFP data sets is presented. A weighted sum of the two bSSFP acquisitions is taken on a voxel-by-voxel basis, effectively synthesizing an off-resonance profile at each voxel that puts fat in the stop band while keeping water in the pass band. The technique exploits the near-sinusoidal shape of the bSSFP off-resonance spectrum for many tissues at large (>50 degrees) flip angles.
Balanced steady-state free precession (bSSFP) is a rapid and SNR-efficient imaging method, but suffers from characteristic bands of signal loss in regions of large field inhomogeneity. Several methods have been developed to reduce the severity of these banding artifacts, typically involving a penalty in scan time. Multiple phase-cycled bSSFP data sets can be acquired and combined in various ways (linear combination, maximum intensity, sum of squares) to reduce the severity of banding, as described in [1-9]. Another technique, wideband SSFP, utilizes two alternating repetition times (TR) with alternating RF phase to widen the band spacing in bSSFP, thereby reducing banding . Slow modulations in the spectral offset frequency have also been used to reduce banding .
Fat suppression with bSSFP is also challenging. Numerous suppression techniques have been proposed, including linear combination , TIDE , FEMR , IDEAL and other Dixon methods [15-17], phase-sensitive fat detection [18,19], alternating TR (ATR) , and periodic fat saturation methods . However, most methods require an additional increase in scan time, and some are incompatible with bSSFP band-reduction techniques . Furthermore, these techniques have varying degrees of robustness to off resonance, and some suffer from partial volume effects [18,19]. Some are transient techniques, relying on a transient magnetization preparation to suppress fat [13,21], while others work in the steady state [12,14,15,19].
This work was motivated by the need for both robust fat suppression and band reduction in the presence of field inhomogeneity when using bSSFP for flow-independent peripheral angiography [22, 23]. Relatively large flip angles (~50 degrees or larger) are desirable when using bSSFP for peripheral angiography to increase vessel conspicuity and improve contrast . This unfortunately can lead to SAR considerations and longer TR. As TR increases, the nulls in the SSFP spectral profile responsible for banding artifacts are more closely spaced in frequency. This can increase the number of signal null bands appearing in the image over a given range of off-resonance, while simultaneously making the transitions from signal to signal null more abrupt spatially; the bands of signal null get thinner and closer together, and consequently the edges get sharper. While the degree of band reduction is unaffected by TR, rapid spatial modulations in signal are typically less desirable than slower spatial modulations. These problems are exacerbated at higher field strengths (e.g., 3 Tesla). Fat suppression can also be compromised for a number of the aforementioned techniques as TR increases [14,20]. Ideally, we would like a bSSFP technique that robustly suppresses fat without partial volume effects or artifacts, reduces bSSFP banding artifacts, works well at large flip angles, works in the presence of large field inhomogeneity, and keeps the scan time penalty to a minimum.
In this work, we present Large-Angle Multiple-Acquisition (LAMA) bSSFP, a novel bSSFP technique that simultaneously suppresses fat and reduces banding artifacts with the acquisition of only two phase-cycled bSSFP data sets. The technique relies on the near-sinusoidal shape of the bSSFP off-resonance spectrum for many tissues at large flip angle, and thus performs best under the large-flip-angle constraint of the flow-independent peripheral angiography application. Preliminary results are presented both in phantoms and in vivo, along with the results of a simulation and study assessing the performance of the technique across a range of flip angles.
As the flip angle is increased, the bSSFP signal as a function of off-resonance frequency begins to approximate a sinusoid for many biological tissues (Figure 1). (Note that this approximation strictly speaking requires echo time TE = TR/2. At other TE, a simple linear phase is introduced across the spectral profile and the signal can be modeled by a complex exponential.) This spectral profile can be arbitrarily shifted in frequency by incrementing the phase of the RF pulse by some Δϕ from excitation to excitation [3,4,7,8]. When two large-flip-angle (~50° or greater) bSSFP acquisitions are performed with Δϕ = 0° and Δϕ =180°, the spectral profiles of the acquisitions approximate a sine and cosine, respectively, with period 2/TR Hz. A data set with spectral profile shifted by a desired spectral shift Δf can then be synthesized from the two acquisitions, using the relationship
as illustrated in Figure 2.
Choice of TR such that TR = n / (2*CSf), where n = [1, 3, 5, …] and CSf is the absolute chemical shift of fat relative to water (in Hz), will place fat in a signal null whenever water is at a signal maximum. Choice of TR to adjust the relative phase of fat and water is also exploited in . If the appropriate frequency shift Δf needed to place water on a signal maximum is then identified on a voxel-by-voxel basis, a fat-suppressed image without banding can theoretically be formed from just two phase-cycled acquisitions. This forms the basis for our large-angle multiple-acquisition (LAMA) bSSFP technique.
The required frequency shift for each voxel can be determined either through acquisition of a field map, or through an intelligent search of voxel intensities over a range of Δf values (eliminating the need for field map acquisition and the associated scan time penalty). In the latter case, a field map is essentially inferred by examining what values of Δf yield the maximum water and minimum fat signal for each voxel. When large jumps in these Δf values are observed in adjacent voxels, the voxels likely lie at a fat/water boundary. Region-growing algorithms can then be devised to identify voxels as predominantly fat or predominantly water, and a smooth map of Δf values synthesized.
Once a smooth Δf map is generated, water-only or fat-only images can be reconstructed. Let S1n be the complex signal of the nth voxel from the first phase-cycled acquisition (with Δϕ = 0°), S2n be the complex signal of the nth voxel from the second phase-cycled acquisition (with Δϕ = 180°), and Δfn be the frequency shift required to put water in the center of the passband (and hence fat in the center of the null) for the nth voxel. The reconstructed signal intensity of the nth voxel in the water-only image (Wn) is then given by:
Similarly, the reconstructed signal intensity of the nth voxel in the fat-only image (Fn) is given by:
The above relations are clearly only valid if the off-resonant spectral profiles of the tissues of interest are identically sinusoidal. While the assumption of sinusoidal spectral profiles is an approximation, we can expect reasonable results across a range of tissues even at flip angles down to approximately 50°. In Figure 3 the spectral profiles of several tissues (fat, arterial blood, venous blood, and muscle) at flip angles of 30°, 60°, and 90° are shown (first column). The second through fourth columns of Figure 3 show spectral profiles synthesized from two phase-cycled acquisitions at Δf values of 1/(12TR), 1/(6TR), and 1/(4TR) respectively. Note that this range of Δf values spans the full range of distortion in the spectral profiles: the distortion at Δf =1/(3TR) is equivalent to that at 1/(6TR), and so forth. While the profiles are distorted enough at a flip angle of 30° to suggest that the LAMA bSSFP method is impractical at these lower flip angles, the distortion is relatively mild at 60°, as illustrated in columns two through four of Figure 3. The distortion continues to decrease for flip angles approaching 180° (irrespective of T1/T2), although signal attenuation becomes significant at flip angles approaching 180°.
The warping of synthesized profiles at different Δf values has an effect on signal contrast. An analysis of arterial/venous, arterial/muscle, and arterial/fat contrast (defined as the absolute signal difference between two tissues) is shown in Figure 4 as a function of both flip angle and resonant shift Δf. As in Figure 3, graphs are shown for Δf = 0, 1/(12TR), 1/(6TR), and 1/(4TR). Interestingly, the distortion tends to improve contrast between the relevant tissues for our flow-independent peripheral angiography application. T1 and T2 values assumed in all simulations for fat, muscle, arterial blood, and venous blood are listed in Table 1.
As an initial proof-of-concept, two phase-cycled bSSFP 3D images and a 3D field map were acquired of an oil/water phantom and the lower leg of a normal volunteer on a 1.5 T GE scanner (GE Healthcare, Waukesha, WI). All subject scans referenced in this work were conducted according to the guidelines of our Institutional Review Board (IRB). Scan parameters for the bSSFP acquisitions were: flip angle = 90° (phantom) and 70° (volunteer), TR/TE = 6.6/3.3 ms, and phase cycling of Δϕ = 0° and 180°. Recall that TR for our technique is constrained to:
where CSf is the absolute chemical shift of fat relative to water (in Hz). Assuming CSf = 226 Hz at 1.5 T, allowable TRs are thus 2.2 ms, 6.6 ms, 11.1 ms, and so forth. A TR of 2.2 ms corresponds to placing fat in the null immediately adjacent to the water passband, and would be expected to be the most robust. However, a TR this short is difficult to achieve. Our choice of TR = 6.6 ms places fat in a null that is one removed from the water passband. A TR of greater than approximately 7 or 8 ms begins to be impractical for SSFP imaging due to the severity of off-resonance banding. As mentioned, increasing TR spatially compresses bands of signal null, leading to more rapid spatial variations in signal and sharper band edges.
For field map reconstruction, a field map was first acquired by taking the phase difference between two successive GRE images with ΔTE = 1/CSf (or 4.4 ms at 1.5T). This ensures that fat and water voxels will be in phase for a voxel that has no off resonance. The GRE images were acquired on the identical matrix as the bSSFP images. The phase difference of each voxel within the image can then be determined to calculate the off-resonance frequency for that voxel within a modulo of 1/ΔTE. The frequency Δfn for the nth voxel was calculated by:
where Δθn is the phase difference in radians of the nth voxel. Once Δf is known for each voxel and both bSSFP images have been acquired, the fat and water images can be created by inserting Δfn into equations (2) and (3).
To eliminate the need for field map acquisition, a region-growing algorithm was implemented that synthesizes a Δf-map using only the two phase-cycled bSSFP acquisitions. This map can be used in place of the field map.
The two constituent phase-cycled images are first processed by creating reconstructed sets of images where each image is shifted by different values of Δf ranging from 0 to 1/TR. (Note that these values only span half of the period of the off-resonance spectra, as we are only concerned with the magnitude of the final image). Afterwards the new set of images are scanned on a voxel-by-voxel basis to find Δfmax for each voxel, where Δfmax is defined as the Δf value that produces the maximum intensity for that voxel, A maximum intensity image is also produced by selecting the maximum intensity found for each voxel across the new set of images. Voxels that are all fat or all water, at the same off-resonant frequency, will have Δfmax values that are separated in frequency by 1/(2TR).
In the peripheral angiography application, the brightest voxels are typically fat. Thus, the assumption is made that the brightest voxel in the maximum intensity image is always going to be a fat voxel with little to no partial volume effects. The algorithm thus assigns the Δfmax of that voxel to be Δffat and Δfwater to be Δffat - 1/(2TR). While this method performs reasonably well for our application, the identification of a fat pixel could be done manually by the MR operator if needed.
Once Δffat and Δfwater for the first voxel are determined, the Δfmax of each neighboring voxel is compared to the first voxel to determine if it is fat or water. If the Δfmax of the neighboring voxel is within a predefined threshold of the surrounding Δffat or Δfwater values, then the new voxel is identified as either fat or water and Δfmax and Δfwater for that voxel are assigned accordingly. If the Δfmax is not close to the known neighboring Δffat or Δfwater values, the voxel is skipped and will not be identified as fat or water.
The algorithm continues in a region-growing fashion until no new neighbors can be identified as either fat or water across the entire dataset. Two slight modifications to the algorithm were found to be helpful in images with large FOV, low resolution, or less than ideal shims. The first modification involves adding magnitude thresholds to limit false identification for voxels close to the noise floor. The second modification is to only use pixels identified as fat for construction of the Δf-map, as these voxels typically have the highest SNR. In some situations, using only fat voxels proved to be more robust in avoiding misidentification in regions of large B0 inhomogeneity.
Once the region-growing algorithm has identified Δfwater and Δffat for any identifiable voxels, a Δffat-map and Δfwater-map are created by interpolating Δfwater and Δffat to fill in voxels that were unidentifiable. When only fat voxels are used, the Δffat-map is created first, and the Δfwater-map synthesized by subtracting 1/(2TR) from the Δffat-map. The Δffat-map and Δfwater-map are then filtered to enforce only gradual off-resonance variation on pixels that were not originally identified. The original images are then reprocessed using the Δfwater-map to make a water-only image and Δffat-map to make a fat-only image.
A simulation was performed to analyze the performance and contrast of LAMA bSSFP at different off-resonance values and at multiple flip angles. A 3D lower leg phantom was simulated in Matlab (The MathWorks, Natick, Massachussetts) by creating a volumetric dataset consisting of a tapered cylinder with fat on the surface and muscle in the middle. Fat pockets of different sizes were randomly placed in the muscle portion. A bifurcated artery and two veins were then added, running superior/inferior (vertically in the images presented). T1 and T2 values assumed for muscle, fat, arterial blood, and venous blood are listed in Table 1. A linear variation in resonant frequency was added to the simulated phantom, also in the vertical (superior/inferior) direction. The off-resonant shift ranged from -1/TR in the most inferior portion of the simulated phantom to 1/TR in the most superior portion (-152 Hz to +152 Hz based on TR = 6.6 ms).
Using this simulated phantom, the expected signal was simulated for two phase-cycled bSSFP acquisitions at flip-angles of 30°, 40°, 50°, 60°, 70°, 80°, and 90°. Fat was assumed to be perfectly off-resonance at 226Hz (the chemical shift at 1.5T). Complex zero-mean Gaussian noise was added with an SNR of 31.6. This data was then processed using the region-growing algorithm as described.
To verify the contrast and performance simulation, 3D LAMA bSSFP images were acquired of a normal volunteer in vivo using a 1.5T GE MRI scanner (GE Healthcare, Waukesha, WI) at flip angles of 30°, 50°, 70°, and 90°. Parameters were chosen assuming a CSf of 226 Hz such that TR = 6.6 ms and TE = 3.3 ms, and two datasets acquired at each flip angle with phase cycling of Δϕ = 0° and 180° respectively. Other parameters were: FOV = 28 × 14 × 14 cm and matrix size = 256 × 64 × 128.
Individual phase-cycled 1.5T acquisitions of both the oil/water phantom (Figure 5(a,b)) and the lower leg of a normal volunteer (Figure 5(g,h)) are shown. Banding is evident in both the phantom and in vivo phase-cycled images. LAMA bSSFP water-only reconstructions using the acquired field maps are shown in Figure 5(c,i), and exhibit relatively good fat suppression in both the phantom and the leg. However, fat suppression appears to be less robust in regions of large field inhomogeneity, such as in the subcutaneous fat. Furthermore, even in regions of relatively homogeneous field, fat suppression is less than optimal, possibly due to noise in the acquired field map.
LAMA bSSFP water-only images reconstructed using the region-growing algorithm are shown in Figure 5(d,j). The region-growing reconstruction effectively suppresses fat in both the phantom and the leg. While better than the field map reconstruction, the region-growing reconstruction is still susceptible to misidentifications of fat in regions of large field inhomogeneity. Note that a small segment of subcutaneous fat was incorrectly identified as water, and hence not suppressed. In regions where voxels were correctly identified as predominantly fat or water, fat/water separation is excellent. For reference, the acquired field maps are shown in Figure 5(e,k), and the maps synthesized using the region-growing algorithm are shown in Figure 5(f,l).
Results of the contrast and performance simulations on the simulated lower leg phantom are shown in Figure 6 for flip angles ranging from 30° to 90°. All of the images are maximum-intensity projections (MIPs) of the reconstructed 3D dataset. The LAMA bSSFP reconstruction using the region-growing algorithm is shown in the lower panes, while a simple root sum-of-squares reconstruction of the two phase-cycled datasets is shown in the upper panes for comparison. The simulated images highlight the improved fat-suppression of LAMA bSSFP at larger flip angles. Smaller flip angles show regions where the arterial/muscle contrast is less than ideal. The signal decreases from many tissues as the flip angle increases. Arterial/venous and arterial/muscle contrast (defined as the absolute signal difference between tissues) is maximized at flip angles of approximately 50-60°. Note that the region-growing algorithm performs very well in this simulation, given the enforced slow variation in field homogeneity.
Results of the contrast and performance verification in vivo at 1.5T are shown in Figure 7 for flip angles of 30°, 50°, 70°, and 90°. As with the simulation results, all images are MIPs through the reconstructed 3D dataset. As expected, some fat-suppression occurs at lower flip angles, but the fat-suppression performance of LAMA bSSFP improves as the flip angle is increased. Residual off-resonance signal modulations (banding) also decrease as the flip angle is increased. The ability of the region-growing algorithm to correctly identify the frequency shift (and consequently to classify voxels as either fat or water) improves as the flip angle increases. Less distortion is also expected in the reconstructed off-resonance profiles, as demonstrated in Figure 3. The in vivo results appear to correspond well with the results expected given the simulations presented in Figure 6. However, focal signal loss that mimics pathology is evident in some areas. We suspect this arises from errors in the synthesis of the Δf-map, but a more thorough exploration of the source of these artifacts is needed.
There are several clear limitations of the LAMA bSSFP technique that should be addressed. First, the requirement of relatively large flip angles (greater than ~50°), while reasonable for the flow-independent peripheral angiography work that motivated this research, could limit the utility of the method for other bSSFP applications. In many situations, lower flip angles are desired to maximize the bSSFP signal. Furthermore, the large flip angle requirement may lead to SAR considerations, particularly at higher field strengths.
The constraints imposed on TR by the LAMA bSSFP technique may also make it impractical for certain applications. As previously mentioned, TR must be chosen such that TR = n / (2*CSf), where n = 1, 3, 5, … and CSf is the absolute chemical shift of fat relative to water (in Hz). In practice, only TRs between approximately 2.5 ms and 10 ms are practical for most bSSFP applications. This limits choice of TR in the above equation to n = 3 for 1.5T and n = 5 for 3T. While we have achieved good results at both 1.5T (with n = 3) and 3T (with n = 5), n values of greater than 5 are likely to result in poor performance for LAMA bSSFP, as both the fat null and the water passband become narrower in frequency and further removed from each other. At 7T, we expect LAMA bSSFP to be largely impractical.
In this work, the coronal orientation was used such that blood flow is mainly along the readout (superior/inferior) direction, thus minimizing the impact of flow-dependent phase variations. Flow-dependent phase variations are expected to have an adverse impact on the performance of the LAMA bSSFP technique, as the technique relies on the predictable variations in signal phase arising from off-resonance. Further work is needed to understand the sensitivity of the technique to flow-induced phase.
As reported in the previous section, we were able to achieve better results with the region-growing reconstruction technique than by using a field map to inform LAMA reconstruction. The region-growing synthesis results in a Δf-map that partially compensates for the distortions caused by the fact that off-resonance spectra are not truly sinusoidal. However, there is still significant room for improvement and optimization of the LAMA technique. We have experimented with several different filtering methods to enforce smoothness in the Δf-maps synthesized by the region-growing reconstruction method. Across regions with small and smooth variations in B0, these filtered Δf-maps yield very good results. However (and not surprisingly), we have found that the region-growing technique achieves better fat suppression in regions of rapid B0 variations when the filtering step is eliminated. Further work is needed to tailor and optimize LAMA bSSFP for various possible applications.
Synthesis of the Δf-map during region-growing reconstruction hinges on the successful classification of a sufficient number of voxels to allow interpolation of the missing points. One reason this classification might fail is simply due to insufficient signal for that voxel. However, failure can also result from partial volume effects (e.g., when a voxel has comparable signal contributions from both fat and water). A region-growing LAMA bSSFP reconstruction can fail if too large a proportion of voxels split their volume across fat and water. We have sometimes observed this failure in muscle tissue with significant fat content. In these cases, a field-map based reconstruction might perform more reliably.
While partial volume effects may cause synthesis of the Δf-map to fail, it is important to point out that the LAMA bSSFP technique is in general robust to partial-volume effects provided an accurate Δf-map is available. If the true off-resonance shift of a given voxel is known, fat signal from that voxel can be separated from water signal by the LAMA reconstruction technique. Thus, while partial-volume effects may result in synthesis of a poor Δf-map using a region-growing technique, they should not adversely affect a LAMA bSSFP reconstruction performed with an accurate Δf-map. However, it is important to note that the relatively broad fat spectrum with multiple peaks will still degrade the performance of the technique. Further work is needed to understand the magnitude of this degradation.
Future work will focus on optimization of the LAMA bSSFP technique for flow-independent peripheral angiography. A more detailed comparison and analysis of the Δf-maps synthesized from region-growing techniques vs. those derived directly from a field map could help inform this optimization. In some cases, the small amount of time needed to perform a low-resolution field map for reconstruction may be a valid trade-off if reconstruction is significantly improved. Hybrid techniques that employ both a field map and a synthesized Δf-map could also be explored to make the technique more robust. Furthermore, a detailed analysis and comparison of the performance of the LAMA technique to other SSFP fat suppression and band reduction techniques needs to be performed.
In summary, large-angle multiple-acquisition (LAMA) bSSFP is a novel technique that simultaneously suppresses fat and reduces bSSFP banding artifacts with the acquisition of only two phase-cycled bSSFP data sets and a field map (in the case of a field-map based reconstruction). Preliminary implementations have been demonstrated with promising results. In addition to the field-map based reconstruction, a region-growing reconstruction that eliminates the need for field-map acquisition has been successfully demonstrated at both field strengths. The technique shows particular promise for flow-independent peripheral angiography using bSSFP, where large flip angles are desirable to achieve good vessel conspicuity. Further work is needed to fully optimize the proposed technique, compare its performance to other bSSFP techniques, and assess its applicability to other potential applications.
This work was supported by NIH 1 R01 HL075803-01, NIH R01 EB002524-01, NIH 5 K08 CA112449, The Ben B. and Iris M. Margolis Foundation, and Brigham Young University.