Magnetic resonance imaging studies of functional brain activation: analysis and interpretation

We have demonstrated that a time series of echoplanar images can contain low frequency noise components which confound analysis of functional MRI data. In simulated tasks of long duration, the false positive rate from t-test analyses greatly exceeded the statistical probability level. As task durations were shortened, the false positive rate declined. We also demonstrated that voxels representing extensive regions of the brain covary significantly over time. This covariation challenges the independence assumption of t-test and other analytical procedures and likely contributes to the false positive rate. The frequency spectra of many voxels showed relatively little power at higher frequencies with the important exception of some blood vessels (Fig. 12). Experimental designs in which stimulus or task conditions were alternated at these higher frequencies (e.g. 0.083 Hz corresponding to a 6 sec task duration and a 12 sec period for a complete two task cycle) did not show an inflated false positive rate when analyzed by t-test. We used the alternating tasks design with task durations of 8.73 sec, 6.4 sec, and 6.0 sec coupled with a frequency domain analysis strategy in a series of somatosensory, motor, perceptual, and working memory experiments. This combination of design and analysis was successful in identifying reliable activations across groups of subjects with a minimum of apparently spurious activations. By introducing a 180 degrees phase shift by reversing task order, we have been able to eliminate the contribution of most high frequency noise sources (such as large blood vessels). By segregating low frequency noise from the frequency of stimulus alternation, we routinely generate stable results in the presence of low frequency noise and drift. Despite the usefulness of the rapid task alternation and frequency domain techniques demonstrated here, there are potential problems and limitations in their application: 1. The short duration of our tasks results in an approximately sinusoidal activation waveform. With longer duration tasks, the activation time course would appear more square with a more complex frequency spectrum than the single peak demonstrated above. In such circumstances we have used convolution analysis with an expected waveform (McCarthy et al. 1996), similar to the approach of Bandettini et al. (1993). 2. If the activation in one task condition is significantly delayed and extends well into the period of the second task, it will be difficult to determine which task produced the activation. This problem is not specific to frequency analysis, and would occur as well for t-tests. One solution we have used is running a single active task against a relatively neutral control such as fixation to determine the usual activation dynamics of the active task. 3. Common activations by two alternating tasks are de-emphasized. This problem is also not specific to frequency analysis, and in most circumstances is an advantage rather than a disadvantage. However, if uncertain as to whether a task is capable of producing any activation, we have again used the strategy of running the task against a relatively neutral control. 4. Some tasks do not lend themselves to the short durations used here. 5. The frequency domain procedures used are conservative and may underestimate the true anatomical extent of the activation. In practice we compute t-tests in addition to the frequency domain techniques to guard against this possibility. Many of the advantages of the procedures described here are due to the alternation of short duration tasks rather than the application of frequency domain techniques per se. However, the success of these techniques in isolating periodic task-related signal changes suggest that a more complex design with concurrent stimulation presented at different frequencies might be feasible. Such designs may have advantages in that categories of stimuli would not be presented in isolation but against a changing ba