False Cerebral Activation on BOLD Functional MR Images: Study of Low-amplitude Motion Weakly Correlated to Stimulus

Methods

Results

The phantom-pneumatic system provided motion control with average in-plane translations (mean ± SD) of 0.74 ± 0.36 mm, through-plane translations of 0.01 ± 0.08 mm, and rotations around the x (pitch), y (roll), and z (yaw) axes of 0.24 ± 0.05°, 0.02 ± 0.06°, and 0.47 ± 0.42°, respectively. For all experiments, in-plane translation and yaw were the dominant movements as intended, showing higher amplitudes and levels of correlation with the boxcar reference than the through-plane parameters. There was generally close agreement between the intended and observed levels of stimulus-motion correlation, with coefficients (r) ranging from 0.31 to 0.96. Control experiments showed negligible displacements that tended to drift linearly with maximum amplitudes near the lower limit detectable by the realignment algorithm (2) (approximately 0.01 mm). Example phantom movements are shown in Figure 3.

The number of voxels subjected to statistical analysis for each phantom image was 20569 ± 78 (mean ± SD), a number comparable with that used in parametric map studies of activation in humans. There were no regions of false activation identified for motions with r < 0.52, at either the low or high Z thresholds (Fig 4). False activations occurred in every experiment with r > 0.67, irrespective of the threshold. The number and sizes of significant regions varied slightly with the threshold (Fig 5), with the expected trend toward fewer and larger regions at the lower threshold and greater numbers of smaller regions at the higher threshold. However, the results remained consistent with respect to the likelihood of observing false activation at a particular level of correlation. Among experiments with r > 0.67, the lower Z threshold resulted in a median of two activated regions per experiment with an average size of 119 voxels (8.4 cc); at the higher threshold, the median was 3.5 regions with an average size of 71 voxels (5.0 cc).

Regions of false activation tended to be located in the vicinity of a boundary between the gelatin and the embedded vials, which corresponded to regions of slightly different magnetic susceptibilities. However, the configurations of the activated areas frequently did not conform to the shape of the boundary, such that they would not be easily recognized as artifact in an actual functional MR imaging study. Several examples of falsely activated regions are shown in Figure 6.

In our four experiments with a resting human participant subjected to passive head movements, displacements were slightly larger than average for the phantom but remained less than ±2 mm/0.5° in-plane and ±0.2 mm/0.3° through-plane. Correlation with the boxcar reference ranged from 0.64 to 0.75. The mean number of voxels subjected to statistical analysis was 19437 (±458 SD), similar to the phantom. Activation was identified in two experiments using the lower Z threshold (one region per experiment; 55 and 90 voxels [3.9 and 6.3 cc, respectively]) and in one experiment using the higher threshold (one region of 24 voxels [1.7 cc]). These regions appeared to be located in cerebral cortex (Fig 7).

Discussion

Several investigators have reported that patient motion during BOLD functional MR imaging cerebral activation studies occasionally produces regions of false activation, especially when these movements are relatively large (>3 mm) and highly correlated with the stimulus (5–8, 14). Conversely, motion artifacts may potentially obscure areas of true activation (5). To overcome these problems, most functional MR imaging studies include some means to limit or detect head motion, such as a bite bar (15), review of images in cine mode (16), and/or image realignment procedures coupled with correlative statistical analyses (1, 2). Our experiments confirmed and expanded these earlier results while raising some potentially troubling issues for the interpretation of individual cerebral activation studies.

First, we have shown that very low amplitude motions (<1 mm of translation or 1 degree rotation in-plane) are capable of causing areas of false activation that survive both image realignment and careful statistical analysis in functional MR imaging experiments. Such small degrees of motion are virtually impossible to prevent, even with highly cooperative participants, notwithstanding the use of bite bars, masks, or other restraining devices. Moreover, brain pulsations secondary to the vascular pulse wave routinely create periodic displacements of intracranial structures of approximately 1 mm or more.

The best current realignment algorithms used to process functional MR imaging data are known to be fallible in correcting for patient motion (5). Although several data processing strategies for overcoming these limitations have been proposed (5, 14), an optimal solution has not yet been found and nearly all published functional MR imaging studies to date do not use additional adjustments to correct for motion. Analyzing the functional MR imaging signal from several falsely activated voxels (Fig 8), we commonly observed a subtle but highly correlated signal shift as a function of displacement that was not removed by our standard image realignment algorithm. Therefore, the areas of false activation seen in many of our data sets can be explained by failure of the algorithm to correct for even submillimeter displacements fully.

Second, we have shown that motions that are only modestly correlated (r > 0.52) with the stimulus can also produce false activation artifacts. To illustrate how such partially correlated motions might occur within the context of a functional MR imaging experiment, we performed numerical simulations of representative participant motions in response to an on-off (boxcar) stimulus paradigm (Fig 9). In this hypothetical three-cycle experiment, we noted that a simple, brief “startle response” with rapid return to baseline position after each stimulus produced a correlation coefficient of only 0.33, inadequate to be manifest as a stimulus-correlated motion artifact in our experiments. A step response or partial response to at least two of the three stimuli, however, produced correlation coefficients of sufficient magnitude (r > 0.52) to generate false activations. A full response with time delay between stimulus and motion (not shown in the diagram) could also produce correlation coefficients in the troubling range of r = 0.5 to 0.9.

Considering the complex temporal dynamics of both cerebral activation and human head motion, one might ask whether the false activations resulting from artificial, square, wavelike phantom motions have relevance to actual functional MR imaging experiments with humans. We think that unless the experimental paradigm and statistical analysis are specifically designed to exploit the differences in temporal dynamics between activation and motion (17), these results are relevant. Motion and activation waveforms need not be idealized square-wave functions for their correlations to be relatively high, as evidenced by our finding false activations even with the reference waveform delayed by 6 s, and temporally smoothed, in the regression analysis. Furthermore, stimulus-motion correlations as high as those in this study are not rare in our experience with human participants, particularly for motor tasks.

Of course, the range of correlations at which false activations become likely will depend on the specific task paradigm and analysis method. Therefore, no simple solution to this partial correlation issue can be stated. The use of “single-trial” or “event-related” paradigms may be useful for distinguishing partially correlated motion effects from brain activation on the basis of temporal dynamics (17). Alternatively, a blocked paradigm may be used as long as the investigator takes care to check the correlation between motion parameters and whatever reference waveform is used in a correlation/regression analysis. The results of our study suggest that the degree of correlation between stimulus and motion is likely to be more important than the amplitude of motion in determining whether a false activation may occur.

Third, we have shown that false activation resulting from stimulus-correlated motion can mimic genuine activation in terms of location and spatial extent. Specifically, we excluded from our results any clusters of falsely activated voxels that were either too small, too large, or too close to the air-phantom boundary to be misinterpreted as real. Although the size and location criteria that we applied were necessarily somewhat arbitrary, it is clear from our results that one cannot assume that false activations will be recognizable as such on activation maps.

The results of our study support and extend the work of others who have investigated functional MR imaging motion artifacts. Hajnal et al (7) found false activations resulting from small displacements in volunteers subjected to visual and motor functional MR imaging paradigms; Wu et al (8) used a motion-control system to produce false activations in a phantom, cadaver brain, and human volunteers. It is important to note that in these previous studies, activation was sought using either subtraction or simple t tests applied to composite “control” and “activation” images. This approach is very different in its sensitivity to motion than more current methods of correlative analysis based on functional MR imaging time series, which are more robust and are preferred to the earlier methods (3). Our study shows that stimulus-correlated motion remains an important problem despite the use of more sophisticated postprocessing.

In all functional MR imaging experiments, there is a trade-off between sensitivity and specificity in detecting cerebral activation. It is possible to remove all motion-correlated signal changes completely by using high levels of filtering and data processing (5, 14). When using these techniques, however, some or all of the true activation being sought will also likely be removed. Moreover, investigators may be justifiably reluctant to apply additional processing to their data when the exact nature of motion effects on functional MR imaging data are not completely understood and there is no consensus regarding how to correct for them.

Whether motion effects will confound data analysis in a particular study will depend on the extent to which a stimulus influences participant motion, the sensitivity of the BOLD signal to whatever motion occurs, and the effectiveness of the realignment procedure in correcting the images. Obviously, stimulus-correlated motion will not pose a problem if it has no appreciable effect on the measured signal or if the alignment procedure adequately corrects for such an effect. Similarly, false activation will not result from a suboptimal alignment procedure if participant movement is not sufficiently correlated with the stimulus.

Finally, we freely acknowledge the numerous methodological limitations inherent in a study such as this. Our phantom provided a well-controlled model that was fairly similar to brain in size, shape, and tissue contrast on functional MR images, with no possibility for true activation to confound the results. On the other hand, the phantom was not identical to brain in its structure or chemical composition. Although no visible distortions at the edges of the embedded vials were observed in the images, the gelatin-plastic-gel interfaces likely introduced microscopic susceptibility distortions potentially responsible for generating motion-correlated artifacts. The relatively higher water content of the phantom as compared with human brain and its different coil loading properties might have rendered a higher signal-to-noise ratio that could make it more sensitive to motion effects than a human participant would be (8).

Although the gelatin and carboxylic salt solutions used in the phantom contained small numbers of macromolecules with chemical shifts different from water, the only resonance peak observed in our experiment came from water protons. No chemical shifts within either compartment could be detected on any images. It seems therefore, that the macromolecules in our phantom were of such low concentration that they were dominated by the water resonance and could not be independently recognized on the overall MR signal. Rather, these macromolecules served to alter indirectly the relaxation times of water protons via cross-relaxation and dipole-dipole interactions, similar to the actions of macromolecules in real human tissues. Such effects have been long recognized based on previous research in similar gel-based tissue phantoms (18, 19).

Our pneumatic motion control system enabled submillimeter, in-plane displacements as determined by the SPM96 realignment algorithm, which is a widely used technique that has been shown to be robust down to 0.01 mm (2). Although the accuracy of this type of image registration in the presence of local task activations has been disputed (20), the phantom had no true activation to compromise the algorithm. It is conceivable that slight motion associated with inflation of the pneumatic system outside the field of view could have perturbed the spatial configuration of the magnetic field to produce correlated signal changes (21); we think, however, that this effect would be negligible compared with that of motion of the phantom itself.

We purposely limited the scope of our investigation to in-plane displacements, which we thought would be less dependent on the specific imaging parameters than would through-plane motion. Because the movements of a human participant will generally include both in- and through-plane displacements, the effects of the latter constitute an important area for future studies. It should also be noted that our conclusions may not be applicable to other pulse sequences, such as 3D or echo-planar acquisitions, which may have different sensitivities to participant movements.

The limited experiments with humans that we included in the study illustrate the problems in performing and interpreting motion effects studies in a living participant. Specifically, in vivo functional MR imaging experiments such as this may contain coexisting areas of true and false activation that would be impossible to differentiate from one another. Regarding human participation in our study (Fig 7), we cannot be certain whether each small area of cortical activation observed was false activation, similar to those generated in the phantom by stimulus-correlated movement, a true resting activation of the brain incidentally correlated with the stimulus, or a true activation in response to sensory stimulus of the air bag. To avoid these complex interpretative issues, we took refuge in a phantom model, which, despite its limitations, could be more carefully controlled.

Conclusion

In summary, participant movements of 1 mm or less that are only modestly correlated with stimuli can produce appreciable false activation artifacts on BOLD functional MR images, even when strict realignment algorithms and statistical techniques are used to exclude them. The degree of correlation between stimulus and motion may be more important than the magnitude of motion in creating these artifacts.

Appendix

Statistical Parametric Mapping Theory

Because of the size and complexity of the data sets generated in BOLD functional MR imaging cerebral activation experiments, sophisticated statistical methods must be used to extract meaningful results. Unfortunately, no consensus has yet been reached concerning the optimal method to detect brain activation. Popular techniques (3, 22–27) include image subtraction, statistical parametric mapping, nonparametric (Kolmogorow-Smirnov) mapping, principal component analysis, fuzzy data clustering, eigenimage analysis, contiguity thresholds, and Fourier transform methods.

Statistical parametric mapping techniques are among the most widely used, and a relatively complete theory of their properties has been developed (12, 13, 28). In this appendix, we define several terms used in our analysis and briefly review the mathematical concepts from which they are derived.

The statistical parametric mapping technique we used was developed by Friston, Worsely, and colleagues in the mid-1990s. It is an extension of 3D data analysis methods used in nuclear medicine and electrophysiology, based on a general block design statistical model known as MANCOVA (multiple analysis of covariance). A powerful statistical principle (the Central Limit Theorem) allows us to transform data sets obtained in functional MR imaging experiments into multidimensional gaussian (normal) distributions. This transformed data (a gaussian field) can then be subjected to tests of statistical inference using Z scores, t tests, F tests, and χ² tests, similar to those used in simple univariate statistics.

To define the global level of statistical significance (P value) for analysis of a single functional MR imaging activation experiment, we define the null hypothesis (H_o) to be “No activation is present.” By setting the global P = .05 in our study, we are stating that the Type I statistical error should be no higher than 5%. In other words, when no true activation is present, we should only mistakenly conclude that a region has been falsely activated by chance less than 5% of the time. (Individual voxels are actually subjected to more stringent P values, generally of approximately P = .001. The global P = .05 figure includes a Bonferroni-like correction for the fact that multiple comparisons of many thousands of voxels are involved in the actual analysis.)

The mathematical analysis defining various conditional probabilities associated with a statistical parametric mapping is extremely complex, involving many terms and variables. Some of the variables (V and W) are defined intrinsically by the functional MR imaging experiment. Other variables (Z) involve somewhat arbitrary decisions by the operator based on perceived levels of clinical significance described below. The basic variables are as follows: V, the total number of voxels analyzed in the sample (approximately 20000 voxels in our phantom); W, a measure of smoothness of the statistical parametric map, calculated by spatial autocorrelation methods from the data; C, the size, in voxels, constituting a region to be tested for significant activation; Z, the minimum acceptable Z score defining threshold activation of a voxel.

As shown by Friston et al (12), the optimum threshold (Z) depends on the shape of the statistical parametric map. As a general rule, broader peaks in a map are best detected by low thresholds and sharp focal peaks are best detected by high thresholds. Based on published recommendations and our own experience, we analyzed each data set using two Z thresholds (Z = 3.09 and Z = 2.56), corresponding to P values of .001 and .005, respectively.

With these terms defined, we can now write an expression for the probability of making a Type I statistical error under the null hypothesis. Specifically, the probability, P, of finding at least one cluster of size C or larger to be falsely activated by chance can be written as follows:where

In these expressions, exp, Γ, and erf stand for the exponential, gamma, and error functions, respectively.