High-resolution (HR) MRI of inhomogeneous tissue environments that have large susceptibility gradients at tissue interfaces is preferably done with spin-echo (SE)-type sequences (1–3). Compared to gradient-echo (GE) sequences, they are less sensitive to signal loss arising from magnetic field inhomogeneity near tissue transitions. On the other hand, practical considerations, such as shorter scan times and higher signal-to-noise ratios (SNRs), make the use of fast GE sequences very attractive for HR-MRI (4–6). Additionally, with recent improvements in gradient hardware, rapid GE sequences with very short TR (TR ≪ T2) and fully balanced gradients over one TR have become very popular for clinical applications. These sequences are commonly known as fully balanced steady-state free precession (bSSFP) sequences. It has been shown that in these sequences the transverse magnetization can also be completely refocused like an SE if the dephasing between excitation pulses remains in the range of ±0.8π (7). However, in the case of trabecular bone imaging (8), a single voxel is composed partly of trabecular bone and partly of bone marrow. The included bone-marrow interfaces give rise to a large range of off-resonances at 3 T that cannot all be refocused. This susceptibility-induced intravoxel dephasing gives rise to artificial broadening of the trabeculae (9). SE-type sequences are less prone to these off-resonance effects because all spins are fully rephased at the echo time (TE).
In order to reduce long scan times, SE imaging usually employs multiple RF SEs to sample multiple phase-encoding lines in one repetition time (TR) (10). However, the long echo train length (ETL) causes T2 blurring, which broadens the point-spread function (PSF) and consequently decreases the effective image resolution (11). Thus, this type of sequence is not suitable for HR-MRI of small structures. Three-dimensional (3D) SE imaging with only one echo per excitation would be optimal. However, the long imaging time required, and thus the low SNR efficiency, limit its clinical use. To address these problems, 3D-SE-type pulse sequences with variable flip angle, such as rapid SE excitation (RASEE) (3, 12) and large-angle SE imaging (13), have been introduced. The main idea is to apply an RF pulse greater than 90°, so that the longitudinal magnetization is partly tipped to the negative axis. The subsequent 180° phase reversal pulse, while generating an echo, partly restores the longitudinal magnetization. Even with this scheme, the TR has to be sufficiently long (on the order of 80 ms (14)) to avoid saturation.
In this work we introduce a new 3D fully balanced steady-state SE (bSSSE) sequence in which all applied gradients are fully rewound over one repetition (Fig. 1). We demonstrate that it outperforms nonbalanced SSSE (nbSSSE) sequences in terms of SNR efficiency (14) while applying a reduced flip angle (<140°). We assessed its performance in in vivo imaging of trabecular bone microstructure in comparison with previously used nbSSSE and fast GE-based pulse sequences. In HR-MRI the fatty bone marrow has a high signal intensity and trabecular bone has a low signal intensity due to the relatively long and short spin-spin relaxation times (T2), respectively (15–17). Thus, in HR-MR images trabecular bone appears dark in between bright streaks of marrow. A major issue in MR-derived visualization and quantification of trabecular structure arises from the fact that the achievable spatial resolution in vivo of MR images is often comparable to the thickness of the trabecular bone itself (80–150 μm) (18). Imaging at higher resolution is primarily hampered by the requirement of long acquisition times due to SNR considerations. At resolutions on the order of the trabecular dimension, partial-volume effects arise due to voxels composed partly of bone and partly of bone-marrow spins. Therefore, MR-derived structural measures, such as trabecular thickness and spacing, are not identical to true histological dimensions (19, 20). However, the trabecular parameters can still be reproducibly measured because they are closely correlated to μCT data (18), and the trabecular spacing (∼800–1000 μm) is usually much higher than the voxel size. The result of insufficient image resolution is not only blurring of the very thin trabeculae but also signal attenuation within the voxels consisting of both bone and bone marrow (21). This is mainly due to the above-mentioned off-resonance effects arising from magnetic field inhomogeneity near the trabecular edges. Additional phase dispersion leading to signal loss is caused by chemical shift in trabecular bone marrow originating from multiple spectral components of triacyl glycerides in fatty bone marrow. Both susceptibility and chemical shift effects lead to a spread in resonance frequency (22). Therefore, an SE sequence with SNR efficiency comparable to fast GE-based sequences would be ideal for imaging the trabecular microarchitecture.
To evaluate the performance of the bSSSE sequence by comparison with previously used sequences for trabecular bone imaging, we implemented an nbSSSE sequence similar to fast large-angle SE (FLASE) (14) and modified it along the lines of Ref.23 to reduce artifacts stemming from spurious stimulated echoes. For the GE-type sequence, a commonly used standard SSFP sequence (5) that acquires the free induction decay (FID) signal immediately after the RF pulse (SSFP-FID), and a bSSFP sequence with multiple acquisitions were considered. In the multiple-acquisition technique based on constructive interference in the steady state (CISS), the resulting images obtained from the acquisitions are generally computed by maximum intensity (MI)-CISS projection (8, 24). The multiple-acquisition method reduces the off-resonance artifacts that are characteristic of bSSFP by giving the magnetization response a more uniform profile (25, 26). The signal of both of the above-mentioned GE sequences is greater than the signal from spoiled GE sequences because the transverse magnetization is maintained and reused after every RF pulse.
The behavior of each sequence was simulated by numerical solution of the Bloch equations (27), and the sequence parameters were optimized for trabecular bone imaging. Experimental studies were conducted on three normal subjects at 3 Tesla (T). The distal radius was chosen as the imaging site because of its high trabecular bone content. After image acquisition, previously described methods (28) were used to compute the apparent trabecular structural parameters, such as the apparent bone-volume over total-volume fraction (app.BV/TV), apparent trabecular separation (app.Tb.Sp), apparent trabecular thickness (app.Tb.Th), and apparent trabecular number (app.Tb.N), for all the three sequences.
Imaging body sites larger than the distal radius, such as the calcaneus, requires a square field of view (FOV) and oversampling in the phase direction, which increases the scan time. In this work we show that even in these cases the imaging time of SE-type sequences can be limited to clinically feasible durations (10–15 min) by the use of a generalized autocalibrating partially parallel acquisition (GRAPPA)-based parallel imaging technique and a reduction factor (R) of 2. Furthermore, we compared the apparent trabecular bone parameters calculated from images acquired using GRAPPA with those calculated from conventional full-FOV images.
MATERIALS AND METHODS
Pulse Sequence Development
We developed a new 3D bSSSE pulse sequence in Environment for Pulse Programming in C (EPIC) for a 3-T Signa system (General Electric, Milwaukee, WI, USA) with a maximum gradient amplitude of 4 G/cm and a slew rate of 15 G/cm/ms. The sequence consisted of a minimum-phase Shinnar-Le Roux excitation pulse and fully rewound gradients (Fig. 1). To avoid a disruption of the steady-state magnetization, all phase encodings and the readout prephasing were performed after the refocusing pulse. Artifacts due to imperfections of the 180° refocusing pulse were eliminated by phase cycling the pulse and placing crushers on either side. We further shortened the TE by combining the crushers with slice-rephaser and slice-encoding gradients. Additionally, an nbSSSE sequence similar to that used in Ref.14, which employs phase-encoding gradients after the inversion pulse, was implemented.
For a fair comparison we optimized all of the sequences for trabecular bone imaging based on simulations of the magnetization response of the sequences by numerical solution of the Bloch equation as described previously (27) in Matlab (MathWorks Inc.). Relaxation times of T1 = 365 ms and T2 = 133 ms for the bone marrow were assumed at 3 T (29). Figure 2 shows the sensitivity to off-resonance frequencies for all sequences. Based on optimization studies for SE sequences, TR was set to 67 ms in both cases and the flip angle was set to 120° (bSSSE) and 140° (nbSSSE) in order to maximize SNR (Fig. 3). A readout bandwidth (BW) of ±8 kHz and a fractional echo (TE = 10 ms) was chosen. The BW was minimized in order to maximize the SNR. Since the radius consists mainly of one type of marrow, and the center frequency was manually adjusted, chemical shift misregistration was not observed in the marrow.
For the GE sequences, a commonly used standard SSFP approach that acquires the FID signal immediately after the RF pulse (SSFP-FID), and a bSSFP sequence (a.k.a. FIESTA, True FISP, Balanced FFE) with two acquisitions (0 and 180 phase-shift increments) based on MI-CISS projection were considered. For the 3D CISS (SSFP-FID respectively) sequence, the readout BW was adopted to ±32 kHz (±16 kHz) in order to minimize the TR according to the limits of gradient heating and RF power deposition. The signal magnitude of CISS is almost independent of TR (for TR < T2) and the theoretically maximum SNR efficiency can be gained by using the shortest TR (8). Short TRs are also necessary to maintain spin-phase coherence. Off-resonance effects disrupt the phase coherence and cause banding artifacts. Thus, the BW that yields the lowest scan time should be used for CISS. A flip angle of 60° (40°), TR = 14 ms (15.8 ms), and TE = 2.9 ms (4.3 ms) were adopted as optimal parameters according to conducted simulations. Furthermore, SNR efficiency (Fig. 4) was calculated for each of the optimized sequences, assuming number of excitations (NEX) = 1.
In three healthy subjects the distal radius was imaged using a GE 3-T scanner (Signa) with the optimized 3D bSSSE, 3D nbSSSE, 3D SSFP-FID, and 3D CISS sequences. Written informed consent was obtained from all subjects and the study was approved by the Committee of Human Research at UCSF. Each subject was positioned in the scanner in a prone position, head first, with the right wrist extended over the head close to the homogeneous center of the magnetic field. The arm was restrained to minimize motion in between scans. The NEX and number of slices (≥32) were adapted for all sequences in order to keep the imaging time below 7 min for all the wrist scans (Fig. 5). Specifically, we used NEX = 1 and 32 slices for bSSSE, NEX = 2 and 38 slices for CISS, and NEX = 4 with 34 slices for SSFP-FID. The signal was detected using a transmit/receive quadrature wrist coil optimized for HR musculoskeletal imaging (Mayo Foundation for Medical Education and Research, Rochester, MN, USA). The image acquisition matrix for all scans was 512 × 384 × 32 using an FOV of 8 cm × 4 cm. Axial images were acquired with an in-plane resolution of 156 μm and a slice thickness of 500 μm.
We additionally acquired images of the calcaneus from four volunteers once using a conventional fully gradient-encoded acquisition and once with GRAPPA with a reduction factor of 2 (30). The scan times for the latter case were 13.7 min and 27.4 min without parallel imaging. The new bSSSE sequence was applied using both the conventional and parallel acquisitions with an eight-channel phased-array head coil (MRI Devices, WI, USA). The calcaneus scans were acquired with a square FOV (10 cm) and oversampling in the phase-encoding direction. The in-plane and through-plane resolution was 195 μm and 500 μm, respectively. The matrix size and all other imaging parameters were equal to the wrist scans. Additionally, we applied the no-phase-wrap option, which effectively samples the phase-encoding direction twice. This is necessary to avoid wrapping artifacts in larger body regions.
Images of the calcaneus were reconstructed offline after the data were transferred to a Sun workstation (Solaris, USA). Images were reconstructed from the fully gradient-encoded acquisition data by conventional Fourier reconstruction. For the partially parallel imaging (PPI) data, a GRAPPA-based parallel reconstruction algorithm developed in our laboratory (30) was applied prior to the Fourier reconstruction. All of the reconstruction routines were programmed in MATLAB (MathWorks).
The subsequent data analysis of the radius and calcaneus was performed using in-house-developed image-analysis software programmed in IDL (RSI, Boulder, CO, USA; and C programming language). A region of interest (ROI) was manually placed by the operator in the distal radius and calcaneus. Ten central sections were analyzed for both sites. The radius was manually outlined by the operator. For the calcaneus, a posterior ROI was chosen to be the largest possible circle that fit the inner perimeter of the posterior portion of the calcaneus (Fig. 6).
The cortical bone was excluded in all cases. Since all four MR sequences were applied consecutively, image registration was not performed and the same ROI positions were used for image analysis of the four sequences. We binarized the ROIs into bone and bone-marrow phases by setting a threshold. This process requires the determination of two reference intensity levels (4) because the voxel size is on the order of the trabecular thickness and there are partial-volume effects. Thus, the histogram shows one single nonsymmetric peak and an asymmetric tail for lower intensities.
We determined the bone intensity IB by sampling the cortical bone intensity at multiple locations through the slices and taking their mean value. The marrow equivalent signal intensity IM was set to the full width at half maximum (FWHM) of the histogram. The apparent trabecular bone volume fraction f = NB/Ntotal was calculated to satisfy the equation f · IB + (1 − f) · IM = I0, where I0 is the mean signal intensity of the ROI to be analyzed (5). The binarizing threshold was then calculated as the intensity value at which the fractional trabecular bone corresponded to the calculated fraction f in the ROI. This resulting threshold was then applied to binarize the MR image. Previously described methods (28) were used to compute the apparent trabecular structural parameters. Furthermore, we determined the SNR by placing circular ROIs in the distal radius. The SNR was calculated as the ratio of the mean intensity (measured in the ROI) to the standard deviation (SD) of the background noise (measured in a region of almost no signal).
Sequence Simulations and Optimization
The resulting transverse magnetization of bSSSE, CISS, and SSFP-FID depends on the distribution of off-resonance frequencies in the voxel, as shown in Fig. 2. The range of off-resonance frequencies in this figure was binned to ±2π (±1/TRbSSFP) for the CISS sequence, since the resulting phase has a period of 4π for TE = TR/2. In the case of a smaller TE, which we consider here, the phase has a larger period. The width of the flat-topped passband of the magnitude is about 35 Hz, and the phase part has a more or less constant value between 10 Hz and 70 Hz in the positive and negative frequency directions (Fig. 2). Note that the magnitude of the SSFP-FID sequence corresponds to the mean signal of an equivalent bSSFP sequence averaged over all off-resonance frequencies, as can be seen in this figure. The phase is linear for the SSFP-FID sequence with a period of ±π (±1/TE), since the spins accumulate a phase due to off-resonance frequencies during TE as previously stated.
A modulation of the amplitude as a function of off-resonances is also visible for bSSSE, similarly to bSSFP, but with a repletion frequency of ± 1/(TR – TE) since the residual transverse magnetization dephases during that time. The width of the flat-topped passband of the magnitude is about 8 Hz and the frequency of these off-resonance modulations is 14.28 Hz (Fig. 2). The amplitudes of the modulations have a magnitude of 0.13. The phases of the spins remain in a range of ±0.5 rad. Therefore, compared to bSSFP, no intravoxel cancellation of the spins occurs and no banding was visible in the acquired images. Additionally, the phase of bSSSE is not opposed to the primary echo, as it is for bSSFP, and therefore no alternating excitation pulse is required (Fig. 2). In contrast, the phase of the nbSSSE sequence is constantly zero at the TE over the entire range of frequencies. Furthermore, simulations showed that the transient response of bSSSE is on the order of T1, similarly to bSSFP. Thus, steady state is reached after approximately T1/TR = 4.5 excitations.
The results of the sequence simulations were used to optimize the sequence parameters. These parameters were then used for MRI of the subjects. Using the optimized scanning parameters, we calculated the SNR for the experimentally used TR, assuming one NEX per sequence. The highest SNR was determined for bSSSE and CISS (0.31), and slightly lower SNRs were found for nbSSSE (0.27) and for SSFP-FID (0.2). SNR efficiency (Fig. 4) was numerically calculated using the optimized scanning parameters and again assuming one NEX per sequence. SNR efficiency was found to be 0.035 (bSSSE), 0.031 (nbSSSE), 0.058 (CISS), and 0.050 (SSFP-FID), and determined as the square root of the TR.
MR images of the wrist using bSSSE, CISS, and SSFP-FID are shown in Fig. 5. The mean SNR from three volunteer scans was higher for bSSSE (13.4) than for nbSSSE (12.9), as expected from the simulations (Fig. 2). SSFP-FID performed similarly to bSSSE (13.4), and CISS had the highest SNR (14.6). The app.TbTh was found to be smaller for all three volunteers using bSSSE compared to nbSSSE. The absence of low-frequency modulations in both the readout and slice-encoding directions (23) demonstrates the successful implementation of the enhanced nbSSSE and subsequently the new bSSSE sequence. The App.TbTh of both GE sequences was significantly higher than those of the SE-type sequences (Table 1). The same is true for app.BV/TV, which was significantly higher using GE sequences and slightly higher for nbSSSE compared to bSSSE. As shown in Fig. 5, the trabecular structure is more emphasized in GE images. SSFP-FID and CISS performed similarly: app.Tb.Th is smaller for CISS in two cases and similar in one case. App.BV/TV was always the highest for SSFP-FID among the four sequences. By applying a parallel imaging technique (GRAPPA), we were able to acquire bSSSE images in a clinically reasonable scan time (∼13 min). The variation in structural bone parameters was below 3.6% for all parameters between the conventional and partially parallel acquisition, which was on the order of the SD (as shown in Table 2).
Table 1. Results from the Experimental Measurement of the Structural Bone Parameters in the Distal Radius (Wrist) are Shown in this Table for Three Volunteers and Each Sequence*
The mean and standard deviations are depicted.
Table 2. Results from the Experimental Measurement with bSSSE of the Structural Bone Parameters Are Shown in This Table for Four Volunteers Using Parallel and Non-Parallel Imaging Technique*
The mean and standard deviations are shown.
In this study we have introduced a new fully balanced SSSE sequence. To evaluate its performance we also implemented an enhanced nonbalanced SE-type sequence used in previous studies (14, 23). We then compared the performance of the bSSSE sequence with a multiple-acquisition bSSFP-type sequence (CISS) and a commonly used GE sequence (SSFP-FID). All sequences were optimized for HR-MRI of the trabecular bone. We presented an analysis of the signal response for all four sequences. The new sequence (bSSSE) revealed higher SNR and SNR efficiency compared to nbSSSE. Additionally, the reduced flip angle allowed shorter TR due to less SAR limitations, and thus shorter scan times. Although there was a modulation of the signal amplitude as a function of off-resonance frequencies for bSSSE (Fig. 2), no banding artifacts were observed in the images. This may be due to smaller modulations compared to bSSFP. However, more importantly, less intravoxel cancellations occurred because the phases of the spins never dephased more than ±0.5 rad (compared to ±pi rad for bSSFP). Furthermore, the T2 value that we used in our simulations (29) may be over-estimated (31), and thus these modulations would be less pronounced.
Both SE sequences performed similarly in terms of depicting the trabecular structure (structural parameters). However, the trabecular thickness was slightly smaller for bSSSE (Table 1). This is due to higher SNR and thus higher contrast between bone and marrow. Although the signal amplitude of bSSSE is a function of off-resonance frequencies (Fig. 2), banding artifacts were not observed in the images (Fig. 5). Also, the fact that the trabecular structure (trabecular thickness) was not enhanced compared to nbSSSE shows that bSSSE is not too sensitive to off-resonance effects stemming from susceptibility differences between bone and bone marrow. This is explained by the phase diagram of Fig. 2. The phases of all spins remain between ±0.5 rad. Thus, intravoxel dephasing, which is responsible for signal loss due to off-resonance effects, is negligible compared to bSSFP sequences (see Fig. 2).
It was previously shown that structural bone parameters derived from both SE- and GE-based sequences have strong correlations with those determined from μCT (18, 22), and with biomechanical properties (32–34). Although the individual trabeculae cannot be resolved accurately, meaningful structural parameters can still be derived due to higher spacing between the trabeculae. We found that GE-type images showed better visualization of the trabecular bone structure. This reflects the enlargement of the trabeculae, as well as enhanced SNR (particularly with the CISS sequence). However, structural parameters differ between the sequences and are the most important parameters for assessing the accuracy of depicting trabecular bone microstructures. From theory and simulations, we found that both SE sequences showed less susceptibility-induced trabecular broadening than the GE sequences. This was expected, since no (nbSSSE) or very little (bSSSE) off-resonance dephasing occurred at the TE using SE sequences, and the resulting image was not distorted by these types of susceptibility artifacts. Using CISS, a complete refocusing of spins was also observed for TE = TR/2 assuming T2 > TR. This refocusing mechanism fails if the magnitude of off-resonance frequencies exceeds ±π which is the case for trabecular bone imaging, especially for voxels that contain more trabeculae than bone marrow. For this reason, the optimum TE depends on the distribution of off-resonance frequencies and thus differs for each voxel depending on its frequency distribution (8). Therefore, TE was optimized for the highest signal and shortest scan time.
The phase behavior of both GE sequences is responsible for the sensitivity to susceptibility artifacts at bone and bone marrow interfaces, since not all spins add constructively to the final signal but can cancel each other out in the worst case. Applying a multiple-acquisition (2) bSSFP sequence (CISS) reduced most of these off-resonance and banding artifacts (25, 26). Although CISS provides more robustness to off-resonance, the signal attenuation inherent in bSSFP type sequences cannot be totally avoided because the complex signal is added for each voxel separately and then reconstructed using the maximum signal intensity for every voxel. Nevertheless, the CISS sequence reveals the highest SNR in a very short scan time despite this signal loss. Additional signal loss is due to T2* decay using GE-type sequences. Signals acquired before and after TR/2 are prone to additional susceptibility sensitivity. In contrast, fast gradient-recalled-echo sequences (SSFP-FID) are not fully balanced (along readout directions) and are thus less sensitive to off-resonance. However, at TE the acquired bone-marrow signal for GE sequences is additionally reduced by a factor of exp(–TE/T2*) due to spin dephasing between excitation and the acquired GE (35, 36). When SE sequences are applied, this additional signal reduction is not observed. This and the non-existence of off-resonance effects are responsible for the relatively high SNR observed despite less signal averaging (NEX). The SSFP-FID sequence shows a more uniformly continuous phase distribution than CISS. However, the phase of the latter sequence is almost constant over a wide range of off-resonances. For larger pixel sizes the range of off-resonance frequencies exceeds 2π in a voxel and thus decreases the signal intensity. That is why for larger pixel sizes the bone parameters shown in Table 1 derive more from the SE sequence but are still similar to SSFP-FID.
The experimental results in Table 1 reveal a strong similarity between both SE sequences and both GE sequences. However, CISS appears to perform better than SSFP-FID, as shown in the table. Setting the correct threshold is much more challenging for MR images (4) than for simulated images because of noise. The better SNR behavior of CISS could therefore have played a major role in binarizing the image more accurately than with SSFP-FID. In addition, the high SNR of the CISS sequence can allow imaging of higher resolution in a short scan time.
A major drawback of using an SE-type sequence is the refocusing hard pulse. The coil geometry is a crucial factor for the quality of the refocusing pulse, since high RF field homogeneity is required. Additionally, scan time is relatively long with SE sequences. Images of the wrist were acquired with an asymmetric FOV to reduce the number of phase encodes to half the number of readout samples. This is not possible for larger regions, such as the calcaneus, where aliasing can be a problem. The phase FOV has to be increased and as a result scan time is doubled and exceeds the clinical reasonable amount. Therefore, we successfully applied a parallel imaging technique (GRAPPA). We were able to reduce the scan time while maintaining similar SNR and structural parameters.
In this work we introduced a new SE sequence for bone imaging. We compared the new sequence with other sequences commonly used for HR-MRI of trabecular bone. We conclude that the new SE sequence yields higher SNR and SNR efficiency with reduced flip angle and TR compared to previously used SE sequences. Compared to GE-type sequences, the new sequence also shows less susceptibility-induced trabecular broadening. In addition, by using a parallel imaging technique based on GRAPPA, we were able to reduce the scan time by half without compromising image quality, and thus established the feasibility of SE-type sequences for trabecular bone imaging in clinical practice.