Real-Time Functional MRI Using Pseudo-Continuous Arterial Spin Labeling

Abstract

The first implementation of real-time acquisition and analysis of Arterial Spin Labeling (ASL) based functional MRI time series is presented in this article. The implementation uses a pseudo-continuous labeling scheme followed by a spiral k-space acquisition trajectory. Real-time reconstruction of the images, preprocessing and regression analysis of the fMRI data were implemented on a laptop computer interfaced with the MRI scanner. The method allows the user to track the current raw data, subtraction images, and the cumulative t-statistic map overlaid on a cumulative subtraction image. The user is also able to track the time course of individual time courses, and interactively select an ROI as a nuisance covariate. The pulse sequence allows the user to adjust acquisition and labeling parameters while observing their effect on the image within two successive TRs. This method is illustrated on a stimulation paradigm consisting of simultaneous finger-tapping and visual stimulation and on a bimanual finger tapping task alternating hands.

Introduction

Conventional functional magnetic resonance imaging (fMRI) collects BOLD-contrast MR images of a subject’s brain while performing a cognitive task, whereas subsequent image reconstruction and analysis are performed offline (i.e., on a separate computer after the experiment is completed). Real-time functional magnetic resonance imaging (rtfMRI) is an exciting extension to conventional fMRI techniques that enables the user to analyze fMRI data as it is being collected. Thus, in rtfMRI the results are immediately available as the subject is being scanned, and can be used to reveal and guide the subject’s cognitive processes. It can also facilitate the experimenter‘s parameter selections or a clinician’s interventions (1).

Several real-time analysis methods have been implemented for online processing of BOLD data, including cumulative correlation (2), sliding-window correlations with reference vector optimization (3), online general linear model [GLM] analysis (4), and combined methods to collect behavioral, physiological and MRI data while performing near real-time statistical analysis (5). All the above methods can facilitate real-time analysis, e.g. the incremental algorithms are useful in monitoring ongoing activation, and the sliding window approaches can improve localization of dynamic activity in time (4). Given current computer processing speed, any of these approaches can be used to display real-time activation maps. This allows for several real-time applications: online data quality control, real-time functional activation monitoring, interactive paradigms based on the subject’s dynamic functional activity (fMRI biofeedback or brain-computer interface[BCI] (6) ), and autonomous control of neural activation using real-time fMRI. These techniques have been used to study subjects’ modulation of motor-area cortical activation and emotional processing (7,8) (9) (10,11).

However, despite the efficiency of BOLD-contrast MR images to detect and localize active site in fMRI, the BOLD signal is difficult to quantify in a physically meaningful way because it results from intricate relationships between cerebral oxygenation, blood flow and volume, as well as the scanner’s unique characteristics. Due to very limited processing time available in rtfMRI applications, quantifying the BOLD signal in real-time poses substantial challenges. Furthermore, the baseline of the BOLD signal drifts over time within a session and this could be quite problematic in experiments where the periodicity of the paradigm is quite long. Lack of meaningful physical units in BOLD data, in addition to these drifts, challenges the use of BOLD imaging for studies of baseline activity and confound long-term activation studies considerably (12).

Given the quantitative nature of its signal and its insensitivity to scanner signal drifts, perfusion based functional MRI using arterial spin labeling (ASL) techniques have emerged as an alternative to BOLD fMRI. ASL techniques (13–15) employ magnetically labeled arterial water as an endogenous tracer to quantify the perfusion rate. This type of measurement is particularly desirable for longitudinal studies and studies with low stimulation frequency, where the BOLD signal drifts are confounded with the signal of interest (16–18). Due to inherently low SNR and CNR in ASL techniques, employing ASL-based FMRI is more challenging than BOLD-based FMRI. SNR maps for BOLD data reported in (19) have values about 50 times higher than those of ASL. Yang et al. (20) also report t-scores of approximately 5 for ASL data versus t-scores greater than 10 for BOLD data. Having said that, the SNR of ASL experiments can be boosted by increasing the length of the stimulation blocks if the cognitive paradigm permits it, as explored at length by Aguirre et al. (21). At the same time, real-time imaging with arterial spin labeling is facilitated by the fact that there is a long period of labeling between image acquisitions.

We note that, in the realm of clinical perfusion imaging with ASL, Xie et al. (22) have developed an adaptive sampling scheme that can determine an optimal set of inversion times (TI) for baseline perfusion quantification with pulsed ASL images. This method examines the images collected and determines the choice of TI for the next image within the acquisition period, i.e., in real-time, but it must be noted that this technique is not related to time series functional MRI.

In this article, we introduce the first implementation, to our knowledge, of real-time, ASL-based, functional MRI. The aims of this communication are to report on the technical feasibility of acquiring perfusion based functional images using ASL in real-time, and the necessary software interfaces and analysis methods for image acquisition and process. Our implementation uses a pseudo-continuous labeling scheme (23) followed by a spiral k-space acquisition trajectory. It allows interactive adjustment of parameters, data smoothing, differencing, statistical analysis and display during acquisition of individual ASL images. It also enables the user to track the time course of individual voxels, and select an ROI as a nuisance covariate interactively. The interface is implemented on a laptop computer interfaced with the MRI scanner.

We demonstrate our implementation of real-time ASL, on a simple visual-motor activation experiment, and a bimanual finger opposition task, although the long-term application motivating this work is fMRI-guided biofeedback to train nicotine-dependent subjects to control their cigarette craving.

Methods

This work was performed on a 3.0T GE Signa scanner (Waukesha, WI) interfaced with an Apple MacBook (Cupertino, CA) with a 2.4 GHz dual core processor and 4GB of RAM. The acquisition pulse sequence was a pCASL sequence (Dai 2009) with (a) phase corrections (19–20) for optimization of labeling efficiency and (b) real-time communication support (c) spiral trajectory image acquisition (FOV = 24 cm, slice thickness = 6 mm, data matrix = 64 × 64 × 8 slices). Each slice generated 4712 complex data points of raw k-space data. The pCASL pulse train had the following parameters: TR=4000 ms, tagging time=2100 ms, post inversion delay=1500 ms, pCASL flip angle=35°, fractional moment=0.9, pCASL TR=1.5 ms, pCASL phase correction=2.2 radians).

A diagram of the data flow is shown in figure 1. Data transfer between the scanner and the MacBook was done through Ethernet using custom C functions modified from the RDS Client software (General Electric Medical Systems, Waukesha, WI). Data were transferred directly from the scanner’s data acquisition board to the laptop’s memory via TCP/IP connection, without disk storage. Reconstruction, processing and display were performed off-line by the MacBook using custom Matlab Software (South Nattick, MA). Transfer and reconstruction were performed after each slice acquisition. To reduce memory requirements, the reconstructed images were analyzed as single precision floating point numbers instead of Matlab’s default double precision. Data were smoothed with a Gaussian kernel, and running subtraction was performed on the images to yield perfusion-weighted images. A reference function reflecting the effects of interest (i.e., the perfusion response to the stimulation paradigm) was constructed prior to the experiment. The reference function consisted of a box-car function convolved with a Gamma-variate hemodynamic impulse response function. This function was used to create a design matrix for GLM analysis that also included a nuisance regressor, which was updated during the data acquisition, and a baseline regressor. As the n^th image was acquired, a temporary design matrix was updated from the first n rows of the whole design matrix. The parameters of this temporary GLM were estimated by ordinary least squares (OLS) and a t-statistic map was computed for the contrast of interest. T-maps were overlaid on the mean perfusion-weighted image and thresholded at a liberal |t| > 3.0. The aforementioned nuisance regressor consisted of the time course at an ROI selected by the user in real-time. This ROI was chosen from a while matter region to remove physiological fluctuations or other spatial correlations in ASL data. All calculations, including image reconstruction were executed within a single TR (4 seconds).

Figure 1. Real-time Processing Stream for ASL FMRI experiment.

Data were transferred directly from the scanner’s data acquisition board to laptop memory without disk storage. Reconstruction, processing and display were performed online by the MacBook using custom Matlab software. Transfer and reconstruction were performed after each slice acquisition. The user interacts with the system by selecting the ROI for the nuisance regressor and a second ROI for time series display.

The global mean time course, the reference function, and the time courses of the user-defined “nuisance region” and a second ROI were displayed in real-time. This allowed the user to monitor activation in any specific brain area and to visually inspect the nuisance regressor for correlations with the reference function. The user was thus able to choose both ROIs in real-time.

For testing and demonstration purposes, two subjects were scanned in accordance with the University of Michigan’s Internal Review Board’s policies, as follows. A real-time pCASL time series was collected while the first subject performed a finger-tapping task consisting of 20 sec tapping with the left hand while a checkerboard pattern flashed in front of her eyes at 8 Hz. The activation period was followed by 30 sec rest. This 50-sec cycle was repeated over a 320 sec period. The second subject was instructed to perform a finger-tapping task with the right hand for 20 sec, followed by 20 sec finger tapping with the left hand. This procedure was repeated for 320 sec. In both cases, the data were transferred, reconstructed and processed, as described above, while the user updated the nuisance ROI, and monitored the image acquisition quality and statistical map in real-time.

Results

After every acquisition, the following displays were updated: (1) current reconstructed image, (2) the most recent perfusion weighted image, (3) the thresholded t-statistic map overlaid on the mean perfusion weighted image, (4) a time course of the spatial mean ASL signal over the field of view, a user-selected ROI, and a second ROI that serves as a nuisance regressor. The model waveform was also displayed for reference. The Matlab reconstruction, processing, and display program was able to keep in real-time (processing lag < 2TR) with the acquisition. Each slice reconstruction from the spiral k-space data took 0.02 sec, on average. The computation time of the OLS over the whole volume increased linearly with the number of time points available, from 0.38 sec at the third acquisition up to 0.57 sec at the 80^th acquisition.

For each subject, a sample screen shot of the user interface taken after 80 acquisitions is displayed in Figures 2 and 3, respectively. Positive t-scores are displayed in a red color scale, while negative ones are displayed on a blue scale. As expected, the visual-motor task (subject 1) produced increases in perfusion that correlated with the reference function in the right motor cortex and the visual cortex. The bimanual motor task (subject 2) resulted in perfusion increases that were correlated with the reference function on the left motor cortex and anti-correlated on the right motor cortex. Figure 4 shows the evolution of the t-statistic map throughout the acquisition for the first subject.

Figure 2. User Interface of Real-time system: visual-motor task.

Screen shot of the user interface during real time acquisition. Clockwise, the panels display the most current unsubtracted image, the most current subtracted image, a plot of relevant time courses, and the most current update of the statistical map overlaid on the mean of the perfusion weighted images. The time courses plotted on the lower right are the reference model function in black, the nuisance time course in green, the spatial mean of the ASL image time course in blue, and the time course of a pixel of interest after nuisance removal in red. Four ROIs used as nuisance covariates are indicated in the top-right panel. These ROIs consisted of a 3 × 3 voxel region sampled at a single slice. The regions were placed on the Posterior Cerebral Artery (PCA), the Anterior Cerebral Artery (ACA), a white mater region (WM) and a heavily vascularized left insula region (Ins).

Figure 3. User Interface of Real-time system: alternating finger-tapping task.

As in figure 2, the raw image, subtracted images, statistical maps and time courses are displayed in real-time. In this case, the activation maps indicated correlation with the reference function in the left motor cortex and anti-correlation on the right motor cortex, corresponding to the alternating finger tapping task.

Figure 4. Evolution of the statistical map during visual-motor stimulation task.

As more data were available, the statistical maps became more detailed with a lower number of false positives and more significant scores in the active regions. The placement of the inversion plane and the imaging slices relative to each other is shown on a sagittal view.

We note that, as a result of spiral acquisition without fat saturation pulses, a circular rim appears around the brain in the raw images shown in figures 2 and 3. This is the off-resonance signal arising from fat and it disappears upon subtraction of the control and tagged images, as indicated in the second and third panels.

The effectiveness of removing unwanted variance changes depending on the user’s choice of ROI. Thus, we explored four ROIs (post hoc) as sources of a nuisance regressor and compared the results in terms of their effect on the t-statistic maps. The four ROIs consisted of the posterior cerebral artery (“PCA”), the anterior cerebral artery (“ACA”), a white matter region (“WM”), and a heavily vascularized area in the left insula (“Ins”). These ROIs are indicated in figure 2A. Figure 5 shows maps of the decrease in residual variance maps due to inclusion of each nuisance regressor in the design matrix. This figure indicates that selecting ROIs in the posterior cerebral artery, the residual variance was reduced most dramatically in the vascular tree. Interestingly, selecting the anterior cerebral artery also reduced the residual variance on a wide region along the frontal edge of the brain. Indeed, high variance along the rim of the brain in FMRI studies is usually indicative of bulk head movement (personal observation). Because the anterior cerebral artery ROI was next to the brain’s edge, it is likely that this timecourse contained a large amount of partial volume changes arising from the subject’s movement. The white matter ROI caused a modest and fairly uniform reduction in residual variance (note the difference in color scales). The left insular region’s effect was also modest, relative to PCA and ACA, but was most noticeable in vascular regions, like the PCA and ACA.

Figure 5. Effects of nuisance regressor on residual variance map.

The maps display the change in the residual variance caused by including the nuisance regressor in the design matrix. The greatest variance reduction was observed when choosing nuisance ROIs in the vasculature.

The effect of nuisance covariates on the t-score distribution can be seen in Figure 6. Not surprisingly, the largest effect was obtained when vascular regions were used as the source of the nuisance covariate. Selecting nuisance regressors from the anterior and posterior cerebral arteries shifted the peak toward zero and reduced the distribution’s width. However, the distribution in larger t-score range remained largely unchanged, suggesting a boost in specificity (reduced the number of false positives), while sensitivity remained largely unchanged (the number of true positives detected remained). The number of activated voxels (T>3) obtained using anterior cerebral artery, white matter, left insula and posterior cerebral artery for are 1091, 1174, 1174 and 924 whereas they were 1173 without use of the nuisance covariate.

Figure 6. Effects of Nuisance regressor on the statistical map’fs distribution.

The distributions of the changes in t-scores over the image produced by the inclusion of nuisance regressors (see figure 5) are displayed as histograms. The broken line corresponds to the t-distribution obtained without the nuisance regressor, while the solid line corresponds to the distribution after including the nuisance regressor. Nuisance regressors obtained from the anterior and posterior cerebral arteries produced a shift of the peak toward zero and a reduction of the distribution’s width. We also note that the t-score distribution in the larger regions was largely unchanged.

The spatial distribution of these differences in the t-score maps due to the nuisance regressors can be seen in Figure 7. We noted that the effects of the white matter and insula regressors were subtler than the arterial ROIs. The white matter ROI, however, boosted t-scores largely in the visual and right-motor cortices and not in other regions. Inclusion of the PCA regressor reduced the t-scores in arterial regions, and boosted the t scores in the visual cortex, but appeared to have little effect in the motor cortex.

Figure 7. Effects of Nuisance regressor on statistical map.

The change in the t-maps produced by the inclusion of nuisance regressors can be seen on this figure. The effect of the white matter and insula regressors was subtler (note the color scales) than the arterial ROIs. The white matter ROI, however, boosted t-scores largely in the visual and right-motor cortices. Inclusion of the PCA regressor reduced t-scores in arterial regions, and boosted t-scores in the visual cortex, but appeared to have little effect in the motor cortex.

Discussion

In this communication, we have demonstrated that real-time arterial spin labeling fMRI reconstructon, display, and analysis can be performed on a personal computer. Our target application is to provide neuroimaging feedback to experimental subjects undergoing stimulation paradigms with very long periodicity (> 45 sec). Having said that, real-time ASL scanning can be quite useful in the fMRI setting for many applications, in addition to our own application. Primarily, it allows the operator to evaluate data quality immediately. This point is important because ASL data have an inherently low signal-to-noise ratio, exacerbated by low inversion efficiency, which can be a problem in certain ASL implementations, such as amplitude modulated control (24). In addition, field inhomogeneities, can compromise the inversion efficiency of some ASL techniques such as pCASL (25–27). Hence, we find it extremely useful to be able to evaluate data quality and make decisions about tuning the labeling parameters immediately. The technique also affords the user the opportunity to quickly evaluate the subject’s performance during the scanning session itself. Furthermore, it allows the investigator to determine the subject’s compliance with the task and/or to make decisions about subsequent stimuli or interactions with the subject.

Real-time imaging with arterial spin labeling is facilitated by the fact that there is a long period of labeling between image acquisitions. This labeling period affords the user ample time for reconstruction, image corrections, and analysis “on the fly”. We are presently developing more sophisticated methods to analyze the data, including pattern recognition methods based on support vector machine classification (28). In our current implementation, the 4 second TR offered ample time to process the data using linear regression, even as the size of the data and design matrices increased. For longer time series or reduced TR times, such as those using Turbo-CASL techniques (29,30), sliding-window or incremental regression techniques (4,31) may be necessary.

As with other ASL fMRI acquisitions, the number of slices that can be acquired is restricted by system acquisition speed. The kinetics of the inversion label are such that the entire volume should be collected within an approximately 500 ms window to ensure that the label has sufficient time to arrive at the tissue but not enough to have washed out completely. While our acquisition scheme used a sequential slice acquisition, 3D imaging schemes for data acquisition may also prove to be beneficial, as demonstrated by (33). Parallel imaging will also increase acquisition speed, although at a penalty of computation time (19) so it may require a faster computing framework. Parallel computing technology and Graphics Processing Units (GPUs) can be used to achieve the necessary computational speed for reconstruction and processing (32).

Using the ROI’s time course as a nuisance covariate improved the analyses. An in-depth study of noise removal by nuisance ROI covariates is outside the scope of this article; however, we note that Wang et al. demonstrated that using the global mean signal as a covariate modestly improved sensitivity by removing spatially coherent noise from the time series, provided the global signal was not correlated with the reference function (34). Behzadi et al. also noted that using time courses from pixels with high variance to construct nuisance covariates had a significant effect for noise removal, and that these high variance ROIs corresponded largely with the arterial vasculature (35). Arterial fluctuations are quite visible to the naked eye in ASL time series data, especially when displayed as a movie. As ASL data are collected and displayed in real-time, the signal intensity in arteries fluctuates with the cardiac cycle, and the subtraction process enhances this fluctuation making the arteries quite evident to the user. These fluctuations are also pervasive on the rest of the image, although to a less dramatic extent, so it is not surprising that including their time course as a nuisance regressor was beneficial for the analysis. Thus, real-time acquisition and processing greatly facilitated the choice of a nuisance regressor. The user can thus pick one such region as a nuisance covariate and immediately observe its effect on the time course of the region of interest as well as on the t-map obtained from the GLM estimation. This approach must be used with care, however, and with some prior knowledge of the expected activation pattern and brain vasculature. In general, standard angiographic techniques are not necessary to identify major arteries, as the oscillations in the ASL signal make them quite apparent to the user. It is important, however, that the user does not select a nuisance region whose activity is task related, as that can result in a loss of sensitivity and detection power. Hence, it is useful to display the reference time course along with the nuisance regressor to identify potential correlations. We also note that including additional nuisance regressors always results in loss of degrees of freedom and, while reducing the residual variance and boosting the t-statistics, they can potentially reduce the statistical significance of the result.

Conclusion

In conclusion, we have demonstrated the first implementation of real-time acquisition and analysis of Arterial Spin Labeling (ASL) based functional MRI time series. All calculations, including image reconstruction were executed within a single TR (4 seconds). Our pulse sequence allows the user to adjust acquisition and labeling parameters while observing their effect on the image within two successive TRs. We were able to achieve the reconstruction and analysis on a laptop computer interfaced with the MRI scanner. By tracking individual time courses, and interactively selecting an ROI as a nuisance covariate, we were able to improve signal detection in real-time.