To compare k-t BLAST (broad-use linear-acquisition speedup technique)/k-t SENSE (sensitivity encoding) with conventional SENSE applied to a simple fMRI paradigm.
To compare k-t BLAST (broad-use linear-acquisition speedup technique)/k-t SENSE (sensitivity encoding) with conventional SENSE applied to a simple fMRI paradigm.
Blood oxygen level-dependent (BOLD) functional magnetic resonance imaging (fMRI) was performed at 3 T using a displaced ultra-fast low-angle refocused echo (UFLARE) pulse sequence with a visual stimulus in a block paradigm. Conventional SENSE and k-t BLAST/k-t SENSE data were acquired. Also, k-t BLAST/k-t SENSE was simulated at different undersampling factors from fully sampled data after removal of lines of k-space data. Analysis was performed using SPM5.
Sensitivity to the BOLD response in k-t BLAST/k-t SENSE was comparable with that of SENSE in images acquired at an undersampling factor of 2.3. Simulated k-t BLAST/k-t SENSE yielded reliable detection of activation-induced BOLD contrast at undersampling factors of 5 or less. Sensitivity increased significantly when training data were included in k-space before Fourier transformation (known as “plug-in”).
k-t BLAST/k-t SENSE performs at least as well as conventional SENSE for BOLD fMRI at a modest undersampling factor. Results suggest that sufficient sensitivity to BOLD contrast may be achievable at higher undersampling factors with k-t BLAST/k-t SENSE than with conventional parallel imaging approaches, offering particular advantages at the highest magnetic field strengths. J. Magn. Reson. Imaging 2010;32:235–241. © 2010 Wiley-Liss, Inc.
THE APPLICATION of parallel imaging to blood oxygen level-dependent (BOLD) functional magnetic resonance imaging (fMRI) has been investigated by numerous researchers in recent years (1–5). The advantages of parallel imaging include reduced imaging time per measurement, leading to greater temporal resolution, which is of particular value for event related fMRI studies; shorter read-out times, which lessen image distortion and blurring and improve the sensitivity to and spatial localization of BOLD signal changes (4); a reduction of the minimum achievable echo time, which allows optimum BOLD contrast; and lower energy deposition (6). These factors are especially important at the highest magnetic field strengths, which offer distinctive advantages for BOLD fMRI (3, 4).
In addition to parallel imaging approaches, which operate in the spatial domain (7, 8), other approaches seek to exploit relationships between data at different timepoints in a series of images (9, 10). A third class of data reduction strategies operate on combined spatio-temporal data, employing prior information (11, 12). One such alternative is k-t BLAST (broad-use linear-acquisition speedup technique) and the related k-t SENSE (sensitivity encoding) technique (11), which exploit spatial (k) and temporal (t) correlations in time series of images to reduce the amount of data that must be collected per timepoint. k-t BLAST and k-t SENSE have been widely applied and optimized for cardiac imaging (13–15), but they also lend themselves to BOLD fMRI (6, 11). Aliasing in undersampled images is resolved by k-t BLAST/k-t SENSE using estimates of the temporal variations in signal intensity from a set of training data and a baseline estimate of the static signal (see Ref.11).
SENSE (7) is one of the most well-established and widely available parallel imaging techniques and results of BOLD fMRI studies using SENSE have been characterized in the literature (1, 2). Therefore, it is of interest to compare any alternative approach with SENSE.
The primary aim of this study was to perform a direct comparison of k-t BLAST/k-t SENSE with conventional SENSE applied to BOLD fMRI, using a simple visual stimulation paradigm. In principle, k-t BLAST/k-t SENSE may be combined with any acquisition, but the current implementation precludes its combination with echo planar imaging (EPI). Largely for this reason displaced ultrafast low-angle refocused echo (UFLARE) (16, 17) was employed. An evaluation of the sensitivity to BOLD fMRI of images reconstructed with k-t BLAST/k-t SENSE at different undersampling factors is presented also.
Normal, healthy adult subjects were imaged after they gave informed consent, according to the requirements of the Local Ethical Review Board. The stimulus was an alternating black and white checkerboard (6.7 Hz), with a black screen as a baseline, presented through video goggles. The stimulus was presented in a typical block design, containing five on/off periods each of 11 scans duration.
Images were acquired at 3 T (Achieva, Philips Healthcare, Best, Netherlands) using an 8-channel radiofrequency (RF) head coil for signal reception, a whole body coil for signal transmission, and a displaced UFLARE (16, 17) pulse sequence with the following parameters: TR/TE = 2800/35 msec; T2* weighting by means of an additional delay τ = 30 msec between RF excitation and the first RF refocusing pulse; six dummy echoes prior to the acquisition train; refocusing flip angle 93° to restrict the specific absorption rate (SAR) to 2 W kg−1; field of view (FOV) 240 × 240 mm2; matrix 1282, partial Fourier 0.7. Images were obtained in three slices of thickness 3.5 mm, separated by gaps of 0.5 mm, positioned parallel to the AC-PC line over the calcarine sulcus.
Functional analysis was performed using SPM5 (www.fil.ion.ucl.ac.uk). Each series of images was realigned and registered to the mean before being smoothed with a Gaussian kernel of full-width at half-maximum (FWHM) 4 mm. Each acquisition was analyzed as a separate session. However, a common mask was used for acquired image sets of an individual subject, and in simulated k-t BLAST/k-t SENSE a common mask was used for all images simulated at the same k-t factor (k-t BLAST/k-t SENSE, with and without plug-in). The masks were generated in ImageJ v. 1.41 (Rasband, National Institutes of Health, Bethesda, MD, http://rsb.info.nih.gov/ij/ 1997–2009) by repeated application of a logical AND function to combine the masks generated for each dataset automatically by SPM5 in a preliminary analysis. The first eight and the final nine scans in each time series were excluded from the analysis to avoid the effects of the temporal discontinuity in the k-t BLAST/k-t SENSE reconstructions (11), giving a total of 94 timepoints for each SPM analysis. Analysis was performed with a significance threshold of P < 0.05 (FWE-corrected) and extent threshold of 20 voxels.
Fully sampled data were acquired in nine subjects. k-space data were removed retrospectively to simulate undersampled datasets, which process is referred to as decimation. Offline reconstruction was performed using k-t BLAST/k-t SENSE to simulate nominal undersampling (k-t) factors of 2, 3, 5, and 8. Each training image comprised 11 lines of k-space in the phase-encoding direction with full sampling in the frequency-encoding direction, resulting in effective undersampling factors of 1.7, 2.4, 3.5, and 4.7. Image reconstruction at each k-t factor was performed twice, once with and once without inclusion of training data in the reconstruction before Fourier transformation (“plug-in” of training data) (14). Images reconstructed from fully sampled datasets using the same offline reconstruction technique were employed as reference images for the analysis.
The simulation reconstructs images with a matrix size that is a multiple of the k-t factor. The number of acquisitions (time points) included in the reconstruction must also be a multiple of the k-t factor. Therefore, images reconstructed at k-t factors of 2, 3, 5, and 8 have matrices of 1282, 1292, 1302, and 1282; and the numbers of scans (of the 111 acquired) are 110, 111, 110, 104, respectively.
Seven of the nine subjects were imaged three times using the same functional paradigm in a single session in a semirandomized order. The three acquisitions were 1) fully sampled, 2) SENSE with an undersampling factor of 2.3, and 3) k-t BLAST/k-t SENSE with a k-t factor of 3 (an effective undersampling factor of 2.3 with 11 lines of k-space per training image). Images were reconstructed using standard scanner software. k-t BLAST/k-t SENSE data were reconstructed four times to produce k-t BLAST/k-t SENSE image sets with and without plug-in (14).
Examples of images from a single subject (Subject 1) acquired using SENSE and k-t BLAST/k-t SENSE are shown in Fig. 1. The fully sampled image is more strongly T2-weighted than the other images, due to the longer echo train of 90, compared with 41 and 30 for the SENSE and k-t BLAST/k-t SENSE acquisitions, respectively. In one subject the video goggles presented the stimulus to the right eye only (Subject 5), but as this was the case for all acquisitions of that subject these data were included in the analysis.
Activation maps generated from simulated k-t BLAST and k-t SENSE images are presented in Fig. 2 (Subject 6).
A summary of results is presented in Fig. 3, where means over all subjects of maximum t-value and total number of voxels, in which activation exceeded the significance threshold, are shown. When no activation was detected at the prescribed significance threshold, both maximum t-value and the number of voxels were recorded as null and included as zeros in the analysis. Minor differences between reference images at different k-t factors may be attributed to the different numbers of scans included in the reconstruction, which must be a multiple of the k-t factor (of 111 images per series, 110, 111, 110, and 104 were included for k-t factors of 2, 3, 5, and 8, respectively).
The number of voxels in clusters of activation appears to depend more strongly on the k-t factor than the maximum t-value (Fig. 3c,d compared with 3a,b). Sensitivity to activation is consistently higher with than without “plug-in,” especially at the highest k-t factor of eight (with k-t BLAST the maximum t-value is 11.23 ± 2.19 compared with 2.80 ± 4.29, the number of activated voxels is 155 ± 114 compared with 15 ± 24; with k-t SENSE the maximum t-value is 9.06 ± 3.63 compared with 7.83 ± 3.49; and the number of activated voxels is 97 ± 63 compared with 48 ± 32). Both the maximum t-value and the number of voxels are significantly lower at a k-t factor of eight without plug-in using k-t BLAST than in reference images or with plug-in (Fig. 3a,c). Intersubject variability of maximum t-value, indicated by the size of the error bars, is lower with plug-in than without at all k-t factors (Fig. 3a,b), although this trend is not apparent in the numbers of voxels in clusters of activation (Fig. 3c,d).
Activation was detected in all subjects using k-t BLAST/k-t SENSE and in six out of seven subjects using SENSE.
Activation maps are shown for all subjects in Fig. 4. Consistent with the findings of Study 1, activation is more significant when training data are included in k-space before Fourier transformation (plug-in) and accordingly a summary of results includes only those with plug-in (Fig. 5). Activation generally achieves greater significance in sets of fully sampled than in sets of undersampled images, although the distribution of clusters of activation is largely independent of acquisition or reconstruction strategy (Fig. 4) and the differences in maximum t-values and numbers of activated voxels between fully sampled and SENSE or k-t BLAST/k-t SENSE data with plug-in are not significant (two-tailed, paired t-tests at P < 0.05). Sensitivity to the BOLD response was assessed in terms of maximum t-values and the total number of activated voxels, as in Study 1. In terms of these measures, sensitivity is broadly similar for conventional SENSE and k-t BLAST/k-t SENSE. Greater intersubject variation is observed in the number of activated voxels than in the maximum t-value (Fig. 5b,a, respectively).
The first direct comparison of k-t BLAST/k-t SENSE with SENSE for BOLD fMRI is presented in this work. The sensitivity to the BOLD response of image series acquired with k-t BLAST/k-t SENSE was found to be similar to that of conventional SENSE at an equivalent undersampling factor (Figs. 4, 5), depending on whether or not training data is included in the reconstructed k-space prior to Fourier transformation into the image space (plug-in (14)) (Fig. 3).
Using the standard, clinical 8-channel head RF receive coil employed for this study at 3 T, a direct comparison of k-t BLAST/k-t SENSE with SENSE at higher k-t factors is not possible due to severe and spatially nonuniform loss of signal-to-noise ratio (SNR) when SENSE is employed. However, analysis of the performance of k-t BLAST/k-t SENSE at higher undersampling (k-t) factors was performed by offline reconstruction of decimated data. This approach has the advantage of avoiding differences due to physiological variations or scanner drift between acquisitions. These simulated data show that the “plug-in” of training data increases sensitivity to activation more significantly at the highest k-t factors than at low k-t factors (Fig. 3), particularly in k-t BLAST. The drawback of the simulation approach is its inability to replicate the benefits of shorter echo trains, namely, reduced blurring and T2 weighting (6).
In SENSE (1/√R) represents the maximum achievable SNR relative to the SNR of fully sampled images, which is equivalent to a g-factor of unity (7). The relative SNR in k-t BLAST and k-t SENSE varies spatially depending on the local temporal frequency of signal changes and the total number of images in the series (nimages) from a maximum of √(nimages/R) in static regions to a minimum of (1/√R) in highly dynamic regions (eg, Ref.6). Hence, the theoretical minimum SNR in k-t BLAST/k-t SENSE corresponds to the theoretical maximum in conventional SENSE and on this basis k-t BLAST/k-t SENSE would be expected to perform at least as well as conventional SENSE. However, at an effective undersampling factor of 2.3 BOLD sensitivity with SENSE or k-t BLAST/k-t SENSE does not decrease in proportion to the theoretical reduction in SNR of (1/√R) (Fig. 5), consistent with reports in the literature on fMRI with conventional parallel imaging (1–3, 5), and explained by the dominance of physiological noise in fMRI data at high static magnetic field strengths (1).
k-t BLAST/k-t SENSE is based on estimates of signal change identified in the temporal frequency distribution of signals in training data, potentially providing a more robust reconstruction than hybrid techniques, which update calibration images employed for standard parallel imaging reconstructions throughout the fMRI time series (such as TSENSE (18)) or those that interpolate in spatial frequency and time (such as k-t GRAPPA (19)). Further improvements in the temporal fidelity of k-t BLAST/k-t SENSE are promised by continuing developments, notably by the k-t PCA approach (20).
The advantages of k-t BLAST/k-t SENSE have to be weighed against the disadvantages of possible temporal and spatial blurring introduced by the reconstruction. Temporal blurring may be diminished by including the training data in the reconstruction before Fourier transformation, which explains the amelioration of sensitivity to activation-induced BOLD signal changes with plug-in, especially using k-t BLAST (Fig. 3a). The additional degrees of freedom provided by the separate signals from individual RF coil elements in the k-t SENSE reconstruction may explain the slightly higher sensitivity without plug-in of k-t SENSE than k-t BLAST at the highest k-t factor of 8 (Fig. 3). However, with plug-in k-t BLAST appears to be less dependent on k-t factor than k-t SENSE (compare Fig. 3a with 3b and 3c with 3d).
The benefit derived from plug-in depends on the frequency spectra of BOLD signals and noise in time and space. If the power of BOLD signal changes were contained exclusively in the region of k-space sampled in the training data, sensitivity to the BOLD response with plug-in would be largely independent of k-t factor, and the technique would approach equivalence to keyhole imaging, with the important difference that the SNR in the outer parts of k-space would be greater in k-t BLAST/k-t SENSE due to the greater number of samples over the whole time course. At the opposite extreme, if the power of the BOLD signals were localized in the outer portion of k-space, which is not contained within the training data, the sensitivity to BOLD signal changes would be independent of the use or not of “plug-in” at a given k-t factor. The results shown in Fig. 3 indicate that the BOLD signal energy is concentrated at the center of k-space. If temporal noise at frequencies around the sampling rate were high and localized at the center of k-space, then replacing the fitted data at the center of k-space with the training data might “reintroduce” noise components that had been effectively filtered out by the k-t BLAST/k-t SENSE reconstruction and the sensitivity to BOLD signal changes would be reduced. There is no evidence for this effect in these fMRI data. However, at higher k-t factors the likelihood that training data better model the true center of k-space than the values fit from the whole time series increases and plug-in has been shown to offer a greater advantage (Fig. 3).
In conclusion, k-t BLAST/k-t SENSE at an effective undersampling factor of 2.3 has been shown to provide equivalent sensitivity to conventional SENSE for the detection of BOLD signal changes at 3 T. For functional paradigms similar to that considered in this study the most robust results are obtained from k-t BLAST reconstruction with “plug-in.”
At k-t factors of five or less k-t BLAST/k-t SENSE offers a means of reducing acquisition times for BOLD fMRI at 3 T without unacceptable loss of sensitivity to BOLD signal changes. Following implementation of k-t BLAST/k-t SENSE combined with EPI this approach promises exciting advantages for whole brain fMRI at high magnetic field strengths.