Frequency Stabilization Using Infinite Impulse Response Filtering for SSFP fmri at 3T

Transcription

1 Magnetic Resonance in Medicine 57: (2007) Frequency Stabilization Using Infinite Impulse Response Filtering for SSFP fmri at 3T Ming-Long Wu, 1 3 Pei-Hsin Wu, 2 Teng-Yi Huang, 1 * Yi-Yu Shih, 2 Ming-Chung Chou, 2,3 Hua-Shan Liu, 2,3 Hsiao-Wen Chung, 2,3 and Cheng-Yu Chen 3 The steady-state free precession (SSFP) method has been shown to exhibit strong potential for distortion-free functional magnetic resonance imaging (fmri). One major challenge of SSFP fmri is that the frequency band corresponding to the highest functional sensitivity is extremely narrow, leading to substantial loss of functional contrast in the presence of magnetic field drifts. In this study we propose a frequency stabilization scheme whereby an RF pulse with small flip angle is applied before each image scan, and the initial phase of the free induction decay (FID) signals is extracted to reflect temporal field drifts. A simple infinite impulse response (IIR) filter is further employed to obtain a low-pass-filtered estimate of the central reference frequency for the upcoming scan. Experimental results suggest that the proposed scheme can stabilize the frequency settings in accordance with field drifts, with oscillation amplitudes of <0.5 Hz. Phantom studies showed that both slow drifts and fast fluctuations were prominently reduced, resulting in less than 5% signal variations. Visual fmri at submillimeter in-plane resolution further demonstrated 15% activation signals that were nicely registered in the microvessels within the sulci. It is concluded that the IIR-filtered frequency stabilization is an effective technique for achieving reliable SSFP fmr images at high field strengths. Magn Reson Med 57: , Wiley-Liss, Inc. Key words: blood oxygenation-sensitive steady state; functional magnetic resonance imaging; balanced steady-state free precession; frequency stabilization; IIR filter 1 Department of Electrical Engineering, National Taiwan University of Science and Technology, Taipei, Taiwan, Republic of China. 2 Department of Electrical Engineering, National Taiwan University, Taipei, Taiwan, Republic of China. 3 Department of Radiology, Tri-Service General Hospital and National Defense Medical Center, Taipei, Taiwan, Republic of China. Grant sponsor: National Science Council; Grant numbers: NSC E ; NSC E *Correspondence to: Teng-Yi Huang, Ph.D., Assistant Professor, Department of Electrical Engineering, National Taiwan University of Science and Technology, Taipei, Taiwan, R.O.C. tyhuang@mail.ntust.edu.tw Received 13 February 2006; revised 4 August 2006; accepted 12 October DOI /mrm Published online in Wiley InterScience ( Wiley-Liss, Inc. 369 Imaging techniques based on blood oxygenation level-dependent (BOLD) contrast (1) have been widely used in brain function studies because they provide wide spatial coverage and good temporal resolution. The functional contrast of BOLD functional MRI (fmri) comes from the signal dephasing caused by the presence of paramagnetic deoxyhemoglobin within the diamagnetic tissue environment. In BOLD fmri with echo-planar imaging (EPI) acquisition, problems with image artifacts have been reported, including geometric distortions, signal dropout at susceptibility boundaries, and T* 2 blurring (2). Although previous studies have discussed ways to avoid these problems, and proposed remedies (3,4), the presence of EPI artifacts could still hamper precise registration of the areas of functional activation. An alternative fmri method using balanced steady-state free precession (bssfp) pulse sequences has been proposed to detect the minute frequency shift caused by changes in the concentration of deoxyhemoglobin (5). Since no geometric distortions are visually noticeable in bssfp images, precise mapping of the functional activation areas with anatomic morphology is potentially achievable. In the original version of this approach using RF pulses at large flip angles (e.g., ), functional contrast is created by the sharp magnitude dip and phase transition band at the central reference frequency (5). Miller et al. (6) modified this method by using small flip angles to increase functional sensitivity, in a method termed blood oxygenation-sensitive steady-state or SSFP fmri. With small flip angles, the signal magnitude across the functional sensitivity band, which also corresponds to sharp phase transition, remains relatively high, as opposed to showing a low signal dip. Because of this property, SSFP imaging is suitable for detecting small functional changes in blood oxygenation with higher sensitivity (6). The feasibility of SSFP fmri has been demonstrated on 1.5T systems; however, a wide application of this technique is currently hampered by some technical requirements. The signal characteristics of SSFP fmri lead to a native sensitivity to field heterogeneity, either spatial or temporal. Spatial field inhomogeneity causes dark bands in balanced SSFP (bssfp) images (7), which previous studies addressed by combining several sets of images acquired at different central reference frequencies to obtain fairly uniform functional signals through the image planes (5,6). In contrast to spatial field inhomogeneity, temporal magnetic field instability gives rise to a drift of the central reference frequency. If the central reference frequency changes in such a way that the oxyhemoglobin and deoxyhemoglobin signals become in phase, the functional signal may be completely lost (8). The causes of temporal fluctuations of the main magnetic field can be systematic (e.g., the gradient system (5,6) used or heating of the passive shims due to shim currents (9)) or physiological (e.g., chest motion due to respiration (2,10,11)). It has to be noted that at higher fields, systematic frequency drifts resulting from heating of high-order shimming can be more prominent (9). In addition, for SSFP fmri, heating of the passive shims due to the extensive use of imaging gradients is especially substantial (5,6,8). Therefore, signal fluctuations caused by frequency drifts on high-field systems can greatly impede the observation of activation signal with SSFP fmri. In this study we propose a frequency stabilization method that is based

2 370 Wu et al. FIG. 1. The signal magnitude and phase in bssfp imaging plotted as a function of off-resonance frequency at flips angles of (a) 20 with RF phase alternation, (b) 20 without RF phase alternation, and (c) 5 without phase alternation. SSFP fmri operates in the condition shown in c, with the central reference frequency corresponding to sharp phase transitions placed between the resonance frequencies of the paramagnetic deoxyhemoglobin and the diamagnetic oxyhemoglobin, resulting in functional contrast shown as the solid line in c. The region within the 3-dB drop from the maximum functional sensitivity in SSFP fmri (shaded area) is within only 7 Hz centered at the central reference frequency. The calculations used T 1 /T /80 ms and TR/TE 8/4 ms. on infinite impulse response (IIR) filtering and yields improvements in both SSFP signal stability and SSFP functional signals at 3.0T. We performed high-resolution SSFP fmri with a long acquisition time (for which field stability is essential) to demonstrate the potential of SSFP fmri after IIR-filtered frequency stabilization. THEORY Blood Oxygenation-Sensitive Steady State The signal intensity and phase angle in bssfp imaging was computed for three typical sets of scanning parameters and plotted in Fig. 1 with T 1 /T /80 ms and TR/TE 8/4 ms as a function of off-resonance frequency. At a fairly large flip angle of 20 and with 180 RF phase alternation, as is often used in clinical bssfp imaging (Fig. 1a), the signal magnitude and phase responses were both flat around the resonance frequency (7). This ensures uniform signal intensity without generating dark bands in the images that would occur at 62.5 Hz. When the phase alternation between RF pulses is removed (Fig. 1b), the magnitude dip and the sharp phase transition become shifted to the resonance frequency. If the central reference frequency is placed at the resonance frequency of water

3 IIR Frequency Stabilization for SSFP fmri 371 FIG. 2. When the results from three separate SSFP experimental trials at 7 Hz, 0 Hz, and 7 Hz from the central reference frequency were combined, the narrow sensitivity band in SSFP fmri shown in Fig. 1c could be extended to about 20 Hz. This scheme is used in our submillimeter SSFP fmri experiments. protons surrounding the paramagnetic deoxyhemoglobin, the frequency shift due to blood oxygenation level alteration can be detected by changes in the signal magnitude and phase (5). To further increase functional sensitivity, the SSFP fmri technique uses an RF pulse train with a reduced flip angle, as shown for the flip angle of 5 in Fig. 1c (6). Note that the sharp phase transition at the central frequency (Fig. 1b) is preserved, while the signal magnitude becomes relatively large. Therefore, if the center reference frequency is placed between the resonance frequencies of the paramagnetic deoxyhemoglobin and the diamagnetic oxyhemoglobin, a nearly out-of-phase behavior can be expected that will cause increased signal change in the presence of blood oxygenation-level alterations (6). Combining Trials at Multiple Reference Frequencies Although better functional sensitivity is achieved with small flip angles, the frequency band of functional sensitivity in SSFP fmri is rather narrow because the theoretical frequency shift between oxy- and deoxyhemoglobin is only 10 Hz at 3.0T (5). Figure 1c plots the computation results of SSFP signal and the resultant functional contrast (defined as the absolute value of vector difference between magnetizations at 5 Hz off-resonance for oxy- and deoxyhemoglobin at 3.0 Tesla) for T 1 /T /80 ms, with scanning parameters used in this study (TR/TE 8/4 ms, 5 flip angle without phase alternation). Note that the narrow functional sensitivity band with a sharp phase transition (marked as the shaded area by a 3-dB drop from the maximum sensitivity) was found to be mostly limited within a 7 Hz-region centered at the reference frequency (i.e., 3.5 Hz). In other words, an inappropriate setting of the reference frequency by 5 10 Hz would lead to reductions of SSFP fmri functional sensitivity by a factor of 2 7. Because the functional contrast in SSFP fmri is sensitive only within a narrow frequency band, previous studies combined several fmri trials acquired at different reference frequencies to extend the effective sensitivity band of the SSFP fmri technique (6,8). In a similar manner, in this study we performed SSFP fmri on healthy subjects at three different off-resonance reference frequencies ( 7 Hz, 0 Hz, and 7 Hz) separately (see Materials and Methods below). These off-resonance frequency values were chosen according to the computation results shown in Fig. 1c. With the imaging results from the three trials combined, the sensitivity band was extended to about 20 Hz as shown in Fig. 2. Frequency Stabilization Although we were able to effectively extend the sensitivity band by combining multiple trials, the individual SSFP fmri experiments were still inherently sensitive to instability of the main magnetic field. This is because if the resonance frequency drifts by more than 10 Hz, the oxyand deoxyhemoglobin will become largely in phase, as indicated in Fig. 1c, leading to a loss of SSFP fmri contrast. Field drift can result from indirect heating of passive shims of the magnet when large currents are applied to the active shim coils (9,12) and the gradients (5,6,8). A main field drift of 0.5 ppm/hr (or 1 Hz/min) was previously observed during an MR spectroscopy (MRS) acquisition on a 3.0T system, and was described as an essential concern for high-quality MRS in vivo on whole-body MRI systems (9). Because of the extensive use of imaging gradients in SSFP fmri, the heating from gradient coils may cause field drifts that are even more prominent than in MRS studies (5,6,8,13). To stabilize the resonance frequency for high-quality MRS in vivo, a previous study (9) inserted a reference scan before each MRS acquisition to detect the frequency drift. The Z0 shim current was subsequently adjusted to compensate for the detected field change. In this study we used a similar method to apply an RF pulse with tiny flip angle (typically 3 ) before each dynamic imaging scan to obtain a short reference signal (Fig. 3). However, instead of adjusting the shim current for field compensation, we used the manufacturer s setting to estimate the initial phase of the excited free induction decay (FID), and adjusted the central reference frequency for the subsequent image acquisition accordingly (11). In other words, we requested the system to track the frequency drift continually instead of correcting for it. IIR Filtering The above-mentioned frequency stabilization method can function successfully if the main magnetic field drifts at a sufficiently low rate, and the tracking of frequency drifts is

4 372 Wu et al. FIG. 3. Schematic drawing of the bssfp pulse sequence, modified with an addition of a small-flip-angle RF pulse preceding image acquisition (N ph stands for the number of phase encodings). The instantaneous central reference frequency is obtained from the initial phase of the FID signals following the RF excitation. The central reference frequency is therefore updated once for every dynamic image. sufficiently accurate. If the field drifts originate mainly from heating effects, the slow-change assumption is likely to be largely valid (9,12,14). Because of the tiny RF pulses applied, however, the limited SNR of the reference signals inevitably adds phase estimation uncertainties that can lead to inaccurate tracking of the frequency drifts. In other words, inaccurate estimation of the field drifts can result in an overcompensation of the instantaneous setting of central reference frequency, causing increased short-term instability. Assuming that the field drift is a slow-changing function, it would therefore be beneficial to apply a lowpass filter such that the uncertainties of frequency estimation due to limited SNR could be smoothed out. Consequently, we adopted a frequency stabilization scheme using the principle of simple IIR filtering (15) for dynamic online adjustment of the central reference frequency, as illustrated below. For the n-th dynamic bssfp scan, the reference frequency f IIR (n) was formulated as: m f IIR n w 0 f TinyRF n w i i 1 m f IIR n i, where w i 1 [1] i 0 where f TinyRF (n) is the central frequency estimated using the FID signal from the tiny RF pulse, f IIR (n-i) is the reference frequency setting for the (n i)-th scan after IIR filtering, w i is the weight of the IIR filter, and m 1 is the step number (order) of the IIR filter. In other words, the reference frequency for the current bssfp scan was set as the weighted average of the previous m central frequency settings and the current frequency estimation from the tiny RF pulse excitation. Therefore, it should be noted that the IIR filter method is not based on the same data acquisition. Instead, the updating of central reference frequency settings has to be executed online to track the field drifts. The weights w i of the IIR filter were chosen to exhibit a low-pass filtering behavior such that the weighted average would reflect the expected slowly drifting magnetic field strength. We chose IIR filtering instead of finite impulse response (FIR) filtering because IIR filters yield a sharper and more selective frequency response than FIR filters of the same order (15). In this study the IIR filter weights were determined empirically. By comparing filter performance using various combinations in phantom experiments, we found that a simple third-order IIR filter with a binomial weighting of (i.e., w 0 1/4, w 1 1/2, w 2 1/4) yielded satisfactory signal stability in the SSFP fmri experiments for the MR system used in this study. Hence this filter was chosen for all subsequent investigations. MATERIALS AND METHODS All of the imaging experiments were performed on a 3.0T whole-body MR system (Philips Achieva, Best, The Netherlands) using an eight-channel head coil, with high-order shimming covering the center FOV in phantom studies, and the occipital lobe in visual SSFP fmri. In all of the fmri trials a 5-Hz checkerboard was used as the visual stimulus. Phantom Experiments Three sets of phantom images (TR/TE/flip angle 8 ms/ 4 ms/5, matrix 64 64, FOV 220 mm; scan time 0.5 s for each time frame) were acquired with 2D bssfp imaging to demonstrate the effectiveness of the frequency stabilization methods. Each set of phantom images contained 100 continuous single-slice scans after exclusion of 10 dummy scans, resulting in a total scan time of approximately 56 s for 110 scans, to observe the influences of field fluctuation on the images due to systematic heating. Acquisitions without frequency stabilization, with manufacturer built-in stabilization, and with IIR-filtered stabilization were performed in three separate runs. The dynamic frequency settings for the three runs were recorded. Following the acquisition in each run, we extracted and compared the first and last images from the 100 dynamic scans to observe the pattern changes in SSFP banding due to field fluctuations over the time course of 51.2 s (100 scans). We evaluated the overall signal stability, including short-term fluctuations on the half-second time scale, on a pixel-by-pixel basis by calculating standard deviation (SD) of signal magnitude throughout the 100 scans to generate the SD maps. In addition, we selected a region of interest (ROI) encompassing voxels near the center of the phantom (i.e., where shimming was optimized) to derive mean signal-time curves of three runs to observe characteristics of signal variations with different stabilization schemes. Human Experiments Eight healthy adults (six males and two females, years old) were recruited for a comparison of the effectiveness of different stabilization methods for in vivo SSFP fmri. For this purpose, three sets of SSFP fmri (TR/TE/

5 IIR Frequency Stabilization for SSFP fmri 373 FIG. 4. The first (top row) and 100th (bottom row) images extracted from three dynamic series of SSFP images acquired (a) without frequency stabilization, (b) with Tiny-RF estimated stabilization, and (c) with IIR-filtered stabilization. Shimming was optimized with respect to the center of the phantom. Without frequency stabilization applied, the banding pattern changed substantially in a (dotted box). With Tiny-RF stabilization (b) and IIR-filter stabilization (c), the signal pattern was largely preserved except for the region marked by the dashed box, suggesting long-term stability of SSFP imaging using frequency stabilization. flip angle 8 ms/4 ms/5, matrix 64 64, four slices, voxel size mm 3 ) were obtained from each subject without frequency stabilization, with manufacturer built-in stabilization, and with IIR-filtered stabilization separately, as in the phantom experiments. The four slices were acquired sequentially so that the steady state could be preserved for all slices. In each fmri trial, time frames were acquired with a block design of either [2 ON, 3 OFF, 15 frames/block] or [4 ON, 5 OFF, 8 frames/ block], 2 s/frame, and full visual field stimulus. Thus the total scan time was about 2.5 min. Written informed consent was obtained from each subject. For this part of the experiments, images were acquired with an on-resonance setting of the central reference frequency. In addition to the low-resolution imaging experiments described above, we also performed high-resolution SSFP fmri on three of our eight subjects using IIR-filtered frequency stabilization. The purpose of this part of the experiments was to demonstrate the advantage of the SSFP technique for high-resolution fmri, where a reduction in partial-volume effects may be beneficial for a potential increase in functional sensitivity (16,17). Since high-resolution SSFP fmri was inevitably accompanied by an increase in the total acquisition time, maintenance of field stability became a critical prerequisite for successful experiments. Each of the three subjects underwent SSFP fmri acquired with matrix (TR/TE/flip angle 8 ms/4 ms/5, two slices, voxel size mm 3, frequency stabilization with IIR filtering) with frequency shifts of 7 Hz, 0 Hz, and 7 Hz as three separate trials. To ensure good image registration, tight foam padding was applied between the subject s head and the volume coil to prevent head motion during high-resolution imaging. A visual stimulus (4 ON, 5 OFF, 8 frames/block, 4 s/frame) was given in 72 dynamic scans. The scan time was nearly 5 min for each trial, resulting in a total scan time of slightly less than 15 min with the three off-resonance shifts. Data Analysis for fmri The SSFP fmri data were analyzed using the SPM5 software package (18) without registration or smoothing. In each trial we performed a t-test with an uncorrected P- value threshold set at the level to identify the activated voxels, and then calculated the normalized mean signal-time curves from the original image data. Activation maps from high-resolution SSFP fmri trials acquired at three different central reference frequencies were combined with maximum intensity projections (MIPs) (6), which were then color-coded and superimposed on T 1 - weighted grayscale images for display. In addition, the original images from the respective OFF and ON blocks were averaged separately such that the level of functional signal activations could be directly observed. We extracted and averaged the last five of eight frames from each OFF or ON block, and thus were able to avoid the transient ascending or descending signals, or possible time delay due to slow hemodynamic response, by dropping the first three time points in each block. In this manner we averaged 25 and 20 images to yield the OFF and ON images, respectively. RESULTS Figure 4 shows the first and last images from three dynamic sets of SSFP fmri phantom images containing 100 scans acquired over a time span of 51.2 s. Results from image acquisition without frequency stabilization, with Tiny-RF estimated stabilization, and with IIR-filtered stabilization are shown in Fig. 4a c, respectively. Note that without frequency stabilization, the dark band is shifted from the bottom part to the central portion in Fig. 4a due to field instability (dotted box). With manufacturer built-in frequency stabilization using Tiny-RF pulses, alterations in the banding pattern are greatly reduced, except for some changes in a central portion (Fig. 4b, dashed box). With IIR-filtered stabilization (Fig. 4c), no obvious change in the banding pattern is observed after the 51.2-s time course. The SD maps representing general signal instability from short-term fluctuations on the.5-s scale to the overall 51.2-s time length are shown in Fig. 5a c for no stabilization, manufacturer Tiny-RF estimated stabilization, and IIR-filtered stabilization, respectively. Figure 5a and b both exhibit SDs as large as 17.2% of the signal level in many

6 374 Wu et al. FIG. 5. The SD maps computed from the 100 SSFP images acquired (a) without frequency stabilization, (b) with Tiny-RF estimated stabilization, and (c) with IIR-filtered stabilization. Signal fluctuations as large as 17.2% were seen in many regions for a and b. IIR-filter stabilization reduced the SD to less than 4.4% in c, suggesting both shortand long-term stability. areas. With IIR-filtered stabilization, the signal SD was substantially reduced to 4.4% or lower (i.e., improvements by a factor of 4). Results from Figs. 4b and 5b for the manufacturer Tiny-RF estimated frequency stabilization method suggest that although this stabilization scheme seemed to improve the overall drifts in central reference frequency over the 51.2-s period, short-term stability from one frame to another was not obtained, resulting in a large signal SD (similar to the level obtained without using any stabilization). The ROI analysis shown in Fig. 6 confirms this inference. The signals as measured from the square box encompassing pixels in Fig. 6a demonstrate that even within the region where shimming was optimized, influences from field drifts were still substantial. For the mean signal-time curves shown in Fig. 6b without frequency stabilization, the signals followed a continuous drop by about 36% through the 100 scans, superimposed by minor fluctuations. The slow baseline drift was successfully eliminated with the Tiny-RF estimated stabilization, but with substantially larger high-frequency signal oscillations, consistent with the findings from Figs. 4b and 5b. With IIR-filtered stabilization, slow baseline drifts and high-frequency oscillations were greatly reduced. Within this limited ROI, the signal SDs of three time curves were 11.88% (without stabilization), 4.41% (with Tiny-RF stabilization), and 1.14% (with IIR-filtered stabilization), respectively. Figure 7 plots the estimated drifts of the central frequency in the phantom experiment without frequency stabilization, together with the reference frequency settings that would have been corrected using Tiny-RF estimated and IIR-filtered stabilization methods. The frequency values were plotted with respect to those of the 10 beginning dummy scans. The resonance frequency without using stabilization (dotted curve) showed an increase of approximately 4 Hz over the entire time course of 51.2 s, which is sufficient to cause a substantial loss of SSFP functional signals. The increasing trend was overlaid by prominent oscillations with an envelope of about 2 Hz. Note, however, that in this experiment the actual resonance frequency values during dynamic SSFP scanning were not known. We recorded these estimated drifts from the initial phase values using Tiny-RF pulse excitations without changing the subsequent settings during image acquisition. As a result, estimation uncertainties could result in inaccuracies of the recordings leading to similar fluctuations in the setting of reference frequencies in Tiny-RF estimated stabilization (dot-dashed curve). In contrast, the IIR-filtered stabilization method adjusted the reference frequency in a smoother way with less than 0.5 Hz fluctuations (solid curve), reflecting only the slow-changing tendency. Figure 8a plots the normalized activation curves obtained from a 30-year-old male subject by low-resolution SSFP fmri under three different stabilization conditions, and their functional maps (activation voxels overlaid on T 1 -weighted images with color-coded T-values) are shown in Fig. 8b. All curves were normalized and are represented as percentage changes with respect to the initial signal intensity. Without frequency stabilization, the activation time curve was distorted and drifted. The expected functional signal increase for the later ON blocks was overwhelmed by the signal drifts and hence is barely discernible. The activation curve with Tiny-RF estimated stabilization showed a visually distinguishable ON-OFF pattern FIG. 6. Temporal behavior of SSFP signals within a ROI (a) over a 51.2-s period before and after frequency stabilization (b). It can be seen that without frequency stabilization, magnetic field drifts resulted in a continuous drifting of the SSFP signals. Tiny-RF stabilization removed the slow drifting, but with substantially increased high-frequency signal fluctuations. Baseline drifting and signal oscillations were both reduced with IIR-filtered frequency stabilization (solid line).

7 IIR Frequency Stabilization for SSFP fmri 375 FIG. 7. Estimated center frequency drifts (with respect to the 10 initial dummy scans) during dynamic SSFP imaging experiments. The dotted curve suggests that the resonance frequency had a drift of approximately 4 Hz/min, which should ideally be tracked by the setting of central reference frequencies and with the high-frequency oscillations discarded. The Tiny-RF stabilization scheme erroneously followed the oscillations. With IIR-filtered stabilization (solid curve), the setting of central frequencies tracked only the slow-drifting component of the dotted curve, with fluctuations largely reduced to less than 0.5 Hz in amplitude. and improved baseline drifts, but with relatively largeamplitude oscillations that could obscure the functional signals. With IIR-filtered stabilization, the oscillations in the activation curve were reduced, yielding a less distorted waveform compared to the other two curves. The functional maps in Fig. 8b show more extensive detection of activations and higher T-values with IIR-filtered stabilization compared to the other two experiments. The findings from the human SSFP fmri experiments were in good agreement with the phantom results. Figure 9 shows the results obtained in a 24-year-old female subject by high-resolution SSFP fmri using IIRfiltered frequency stabilization with off-resonance frequencies of 7 Hz (top), 0 Hz (middle), and 7 Hz (bottom), respectively. The two rows in each off-resonance frequency correspond to two different slice locations. In all FIG. 8. a: Activation time curves from the lowresolution visual SSFP fmri experiments using a 4-ON/5-OFF block paradigm. The time span was about 2.5 min. Without frequency stabilization (dotted curve), the drift in signals led to barely discernible functional activations for the fourth ON block. Tiny-RF stabilization was able to restore the signal baseline (dash-dotted curve), but showed relatively large signal fluctuations. IIR-filtered stabilization further improved the activation time curves (solid curve). b: Functional maps showed more spatially extensive detection of activations and higher T-values with IIR-filtered stabilization compared to the other two experiments.

8 376 Wu et al. FIG. 9. High-resolution SSFP fmri experiments using a 5-Hz checkerboard visual stimulus on a 24-year-old female subject, with IIR-filtered frequency stabilization. The images correspond to three separate trials using different settings of off-resonance frequency at 7 Hz (top), 0 Hz (middle), and 7 Hz (bottom). Two image slices are shown for each trial. The OFF and ON images are the averaged results from individual time frames, whereas the functional activation maps were obtained using SPM5 with an uncorrected P-value threshold set at the level and then color-coded and superimposed on T 1 -weighted images. No postprocessing was applied for spatial coregistration. Note that some of the differences in signals between the ON and OFF images are visually distinguishable (small black arrows). trials, we were able to directly observe the functional signal changes from neuronal activation by visually comparing the changes in brightness in the average images from OFF (total 25 frames) and ON (total 20 frames) blocks (small black arrows in Fig. 9). Regions that showed signal increases during ON blocks exhibited good agreement with the activation maps analyzed by SPM5. It can be seen that the activation maps at the three off-resonance frequencies are largely consistent, at least within this relatively small occipital region. The activation maps shown in Fig. 10a were obtained by the MIP combination of the activation maps in Fig. 9 from three off-resonance frequency trials. Most activated voxels were localized nicely near the gray-matter border areas as compared to the underlying T 1 -weighted images. In fact, the activations were mostly concentrated on the microvessels within the sulci, in agreement with previous highresolution fmri reports based on conventional BOLD contrast (17,19,20). The activation curves for the three trials are plotted in Fig. 10b, represented as the percentage changes of the original signal intensities within the activation areas normalized with respect to initial baseline. With IIR-filtered frequency stabilization, no prominent baseline drift was observed in any of these 5-min trials. In FIG. 10. a: MIP combination of the three activation maps in Fig. 9, suggesting functional activation regions located on the microvessels in the sulci, consistent with previous high-resolution fmri findings. b: The activation signals for the high-resolution experiments reached the level of 15% for all trials, reflecting the effectiveness of partial-volume reduction by high-resolution SSFP fmri with IIR-filtered frequency stabilization.

9 IIR Frequency Stabilization for SSFP fmri 377 addition, due to the reduced partial-volume effect in the high-resolution experiments, the activation signals were observed to be as high as 15%, in agreement with the visually appreciable signal changes shown in Fig. 9. DISCUSSION The functional contrast mechanism of BOLD fmri has been relatively well studied (1,21), and this technique provides an important tool for detecting neuronal activations. For fmri with an image matrix of 64 64, EPI acquisitions can yield good temporal resolution and spatial coverage with image artifacts that are often tolerable. With an increasing image matrix, however, the echo train becomes longer and the image quality is progressively affected by T* 2 -related point-spread-function (PSF) blurring (2). Particularly at high field strengths, the EPI geometric distortions that result from increased susceptibility effects are often unlikely to be negligible (22). Remedies such as segmented EPI to alleviate image artifacts (23) or correction schemes with field map acquisition (3) are therefore highly desirable at high field strengths (4). The SSFP fmri technique based on bssfp imaging is potentially a valuable option for distortion-free imaging at a low-rf specific absorption rate (SAR), which is particularly suitable for high-resolution fmri at high fields (6). However, it has to be noted that the range of functional sensitivity for SSFP fmri is highly peaked at the central reference frequency, which consequently makes the SSFP sequence more demanding in terms of field stability than other techniques. In fact, previous studies that demonstrated the feasibility of SSFP fmri on 1.5T systems also reported problems related to frequency drift (8). However, according to our own experience at 1.5T, SSFP fmri can be similarly performed with fair reproducibility on a 1.5T scanner without frequency stabilization. The need to maintain field stability for successful SSFP fmri becomes particularly important at high fields, such as 3.0T, where field drift can be prominent and cannot be neglected. In our phantom study we observed frequency drift at the level of 4 Hz/min on our 3T system. This is four times higher than what was reported for in vivo MRS (9), presumably due to the extensive use of gradient switching in SSFP imaging (8,13). Without frequency stabilization, therefore, functional activation in SSFP fmri can easily be overwhelmed by the presence of large baseline drifting even with a small matrix size of (8). Frequency stabilization implemented by inserting a tiny RF pulse before each dynamic scan should theoretically be able to track the drifting of resonance frequency reasonably well. Due to the limited SNR of the FID signal excited from tiny RF pulses, however, frequency adjustment was seen to be modulated by a fast oscillation of about 2 Hz in envelope, instead of the expected slow-changing behavior of field drifts (9,12). Frequency deviations on the order of 2 Hz would not disturb most MR applications, including spatial displacements along the phase-encoding direction in EPI. For SSFP fmri, on the other hand, since the signal is extremely sensitive to frequency changes, adjusting the central reference frequency in a smooth manner at accuracies better than the 1-Hz level (equivalent to ppm) becomes obligatory for practically feasible SSFP fmri at high field strengths. We think that this is the main reason why the IIR filter setting is the critical factor for the success of SSFP fmri showing a virtually flat signal baseline with fluctuations of 5%. As a side note, the method proposed in our study is not sequence-specific. Therefore, although this article mainly addresses improvements to the low-flip-angle SSFP fmri method (6,8), the same scheme should work well for high-flip-angle SSFP fmri (5), which also requires high stability for the central reference frequency. The parameter selection for IIR filtering could actually be more flexible than the single empirical combination chosen in this study. In fact, the theories of linear prediction state that for time-series data x(n) exhibiting certain autoregressive properties, the forward prediction of x(n) using a linear combination of m previous samples x(n m)...x(n 1) has an optimal solution (15). For SSFP fmri, however, what needs to be predicted is the instantaneous value of the resonance frequency during bssfp scanning, while information about frequency drifting can only be estimated using the noisy FID excited by a tiny RF pulse. Strictly speaking, the intervals between previous samples were not a constant value. Therefore, the theories of linear prediction do not apply to our case of IIR frequency stabilization. As a consequence, our preliminary experiments using digital filters designed from optimal linear prediction did not yield satisfactory results (data not shown). Instead, the simple third-order filter was chosen. When the filter order was increased, the effects on the inhibition of signal oscillations became stronger. However, the stronger smoothing effects were accompanied by a sluggish frequency update that could result in a failure of frequency tracking. In other words, the actual frequency stabilization strategy should also be optimized according to the drifting properties of individual systems. Our proposed scheme of frequency stabilization using IIR low-pass filtering works well only if the assumption of slow-drifting magnetic field holds true. The slow-drift assumption is well supported by evidence from the literature (9,12 14,24,25). To the best of our knowledge, all published reports on magnetic field drifts state that the change occurs rather slowly compared to a time scale of 1 s, whether the magnetic field strength drift is due to mechanical vibrations of the gradients (13), heating of passive shims (9,12), or unknown causes. In our study the actual sources of the field drifts were uncertain, which nevertheless did not hamper the effectiveness of IIR-filtered frequency stabilization. The only exception is respirationinduced field drifts, for which a remedy closely related to the frequency stabilization scheme used in our study was recently reported (11). One major difference between these two techniques is that Lee et al. (11) updated the central reference frequency in every TR, whereas we update the frequency settings in every dynamic imaging frame. As a result of its much higher sampling rate, the method used by Lee et al. (11) is capable of tracking faster changes, such as respiration-induced field drifts, provided that real-time updates of the scanning parameters are readily achievable. In this case, an appropriate IIR filter design may allow a stabilization scheme to simultaneously address both the fast respiration-related drifts and the slow heating-induced drifts. On the other hand, our approach is advanta-

10 378 Wu et al. geous in that it has a simpler design and causes no disturbance to the steady-state signals within individual image acquisitions due to insertion of the extra RF pulse. In terms of combining multiple fmri runs acquired at different resonance frequencies, previous studies discussed methods to combine multiple activation maps, such as taking the maximum statistics (6) vs. taking their mean value (8). In SSFP fmri without frequency stabilization, the frequency drifts during a dynamic SSFP fmri experiment are likely to cause a corresponding spatial shifting in the functional sensitivity region. The resulting signal-time activation curves may exhibit only partial correlation with the stimulus paradigm (see the dotted curve in Fig. 8) in one frequency and a complementary partial correlation in another. By combining multiple trials, one can ensure that the local frequency in the ROI (the visual cortex in this study) is within the extended sensitivity band. In situations involving frequency drift, taking the average maps seems to be appropriate (8). On the other hand, with our IIR-filtered stabilization method, slow baseline drift and fast signal oscillation were both reduced substantially. Regions of functional sensitivity are therefore unlikely to shift spatially during the SSFP fmri experiments. In other words, a given voxel is likely to exhibit a peaked functional contrast in exactly one run at one specific frequency setting, but not the other. Hence, the MIP method seems to be more suitable for SSFP fmri with IIR-filtered frequency stabilization than the average maps, although in our cases these two operations did not exhibit a visually prominent difference. With MIP of activation maps, it makes the best use of the sensitivity plateau on the extended band, as shown in Fig. 2. Potential development options for multiple-frequency SSFP fmri include the acquisition of dynamic scans at interleaved frequencies (e.g., 7 Hz, 0 Hz, 7 Hz, 7 Hz, 0 Hz, 7 Hz,... ). With this method, data at all frequencies can be acquired in a single trial with the subject in the same physiological condition. Furthermore, spatial registration of images acquired at different frequencies is less of an issue compared to acquisitions completed in multiple trials. SSFP fmri is particularly advantageous for high-resolution applications. As demonstrated in this study, a reduction in voxel volume from mm 3 to mm 3 (i.e., a factor of 27) resulted in an increase of the functional activation signal from 5% to 15% despite the SNR decrease in the original images. This result is consistent with reduced partial-volume effects at a smaller voxel size, as has been shown in previous high-resolution fmri studies employing fast low-angle shot (FLASH)-type gradient-echo acquisitions (16,17) or segmented EPI techniques (20). Compared to FLASH-type gradient-echo for fmri, the SSFP fmri technique is superior in its SNR efficiency (26), and the functional contrast can be obtained at short TE and TR (6,8). This leads to substantially less scan time or equivalently a higher efficiency. The lower temporal resolution of SSFP fmri as compared to routine single-shot EPI fmri could be partially overcome by the use of parallel imaging acceleration (27). However, the potential influence of this technique on transient-state vs. steady-state behavior for the functional signals remains to be investigated (28,29). With influence from temporal field drifts appropriately reduced, the static spatial magnetic field inhomogeneities remain the only major obstacle to be overcome before high-resolution SSFP fmri can find wide application (8). CONCLUSIONS We conclude that the SSFP technique is suitable for fmri at high field strengths, provided that a suitable frequency stabilization scheme is employed. The results from this study demonstrate that low-pass IIR filtering can successfully track slow-drifting central reference frequencies with less than 0.5-Hz fluctuations, and thus substantially increase the reliability of SSFP fmri experiments. Our experiments achieved 15% functional signal sensitivity with activations concentrated in the microvessels within the sulci, which suggests that SSFP fmri with IIR-filtered frequency stabilization has strong potential for high-resolution, distortion-free fmri studies. REFERENCES 1. Kwong KK, Belliveau JW, Chesler DA, Goldberg IE, Weisskoff RM, Poncelet BP, Kennedy DN, Hoppel BE, Cohen MS, Turner R, Cheryl H, Brady TJ, Rosen BR. Dynamic magnetic resonance imaging of human brain activity during primary sensory stimulation. Proc Natl Acad Sci USA 1992;89: Jezzard P, Clare S. Sources of distortion in functional MRI data. Hum Brain Mapp 1999;8: Chen NK, Wyrwicz AM. Optimized distortion correction technique for echo planar imaging. Magn Reson Med 2001;45: Ardekani S, Sinha U. Geometric distortion correction of high-resolution 3 T diffusion tensor brain images. Magn Reson Med 2005;54: Scheffler K, Seifritz E, Bilecen D, Venkatesan R, Hennig J, Deimling M, Haacke EM. Detection of BOLD changes by means of a frequencysensitive truefisp technique: preliminary results. NMR Biomed 2001; 14: Miller KL, Hargreaves BA, Lee J, Ress D, decharms RC, Pauly JM. Functional brain imaging using a blood oxygenation sensitive steady state. Magn Reson Med 2003;50: Haacke EM, Wielopolski PA, Tkach JA, Modic MT. Steady-state free precession imaging in the presence of motion: application for improved visualization of the cerebrospinal fluid. Radiology 1990;175: Miller KL, Smith SM, Jezzard P, Pauly JM. High-resolution FMRI at 1.5T using balanced SSFP. Magn Reson Med 2006;55: Henry PG, van de Moortele PF, Giacomini E, Nauerth A, Bloch G. Field-frequency locked in vivo proton MRS on a whole-body spectrometer. Magn Reson Med 1999;42: Pfeuffer J, Van de Moortele PF, Ugurbil K, Hu X, Glover GH. Correction of physiologically induced global off-resonance effects in dynamic echo-planar and spiral functional imaging. Magn Reson Med 2002;47: Lee J, Santos JM, Conolly SM, Miller KL, Hargreaves BA, Pauly JM. Respiration-induced B0 field fluctuation compensation in balanced SSFP: real-time approach for transition-band SSFP fmri. Magn Reson Med 2006;55: Ebel A, Maudsley AA. Detection and correction of frequency instabilities for volumetric 1H echo-planar spectroscopic imaging. Magn Reson Med 2005;53: Foerster BU, Tomasi D, Caparelli EC. Magnetic field shift due to mechanical vibration in functional magnetic resonance imaging. Magn Reson Med 2005;54: Ward HA, Riederer SJ, Jack Jr CR. Real-time autoshimming for echo planar timecourse imaging. Magn Reson Med 2002;48: Haykin SS. Adaptive filter theory. 4th ed. Upper Saddle River, NJ: Prentice Hall; Hoogenraad FG, Hofman MB, Pouwels PJ, Reichenbach JR, Rombouts SA, Haacke EM. Sub-millimeter fmri at 1.5 Tesla: correlation of high resolution with low resolution measurements. J Magn Reson Imaging 1999;9:

11 IIR Frequency Stabilization for SSFP fmri Frahm J, Merboldt KD, Hanicke W. Functional MRI of human brain activation at high spatial resolution. Magn Reson Med 1993;29: Friston KJ, Jezzard P, Turner R. Analysis of functional MRI time-series. Hum Brain Mapp 1994;2: Jesmanowicz A, Bandettini PA, Hyde JS. Single-shot half k-space highresolution gradient-recalled EPI for fmri at 3 Tesla. Magn Reson Med 1998;40: Thulborn KR, Chang SY, Shen GX, Voyvodic JT. High-resolution echoplanar fmri of human visual cortex at 3.0 tesla. NMR Biomed 1997;10: Hennig J, Speck O, Koch MA, Weiller C. Functional magnetic resonance imaging: a review of methodological aspects and clinical applications. J Magn Reson Imaging 2003;18: Yang QX, Wang J, Smith MB, Meadowcroft M, Sun X, Eslinger PJ, Golay X. Reduction of magnetic field inhomogeneity artifacts in echo planar imaging with SENSE and GESEPI at high field. Magn Reson Med 2004;52: Li Z, Wu G, Zhao X, Luo F, Li SJ. Multiecho segmented EPI with z-shimmed background gradient compensation (MESBAC) pulse sequence for fmri. Magn Reson Med 2002;48: Schmidt O, Widmaier S, Bunse M, Jung WI, Dietze GJ, Lutz O. Artifacts in CSI-measurements caused by the drift of the static magnetic field. MAGMA 2000;10: Durand E, van de Moortele PF, Pachot-Clouard M, Le Bihan D. Artifact due to B(0) fluctuations in fmri: correction using the k-space central line. Magn Reson Med 2001;46: Scheffler K, Lehnhardt S. Principles and applications of balanced SSFP techniques. Eur Radiol 2003;13: Pruessmann KP, Weiger M, Scheidegger MB, Boesiger P. SENSE: sensitivity encoding for fast MRI. Magn Reson Med 1999;42: Huang TY, Huang IJ, Chen CY, Scheffler K, Chung HW, Cheng HC. Are TrueFISP images T2/T1-weighted? Magn Reson Med 2002;48: Schmitt P, Griswold MA, Gulani V, Haase A, Flentje M, Jakob PM. A simple geometrical description of the TrueFISP ideal transient and steady-state signal. Magn Reson Med 2006;55: