|Home | About | Journals | Submit | Contact Us | Français|
When events occur spontaneously during the acquisition of a series of images, traditional modeling methods for detecting functional MRI activation detection cannot be employed. The two-dimensional Temporal Clustering Algorithm, 2dTCA, has been shown to accurately detect random, transient activations in computer simulations without the use of known event timings. In this study we applied the 2dTCA technique to detect the timings and spatial locations of sparse, irregular, transient activations of the visual, auditory and motor cortices in 12 normal controls. Experiments with one and two independent types of stimuli were employed. Event-related activation using known timing was compared to event-related activation using 2dTCA detected timing in individuals and across groups. The 2dTCA algorithm detected the activation from all presented stimuli in every subject. When compared to block-design results using a measure of correlation between activation maps, no significant difference was found between the 2dTCA activation maps and the event-related maps using known timing across all subjects. Therefore, 2dTCA has the potential to be an accurate and more practical method for detection of spontaneous, transient events using fMRI.
Functional MRI (fMRI) is a powerful tool for localizing activity in the brain in response to a controlled stimulus. The majority of fMRI studies rely on the use of known stimulus timings to detect voxels whose signal intensity varies with these timings. However, there are many applications where fMRI signal changes are expected but they occur spontaneously and cannot be induced in a controlled manner. These include interictal activity in epilepsy, tics in Tourette's Syndrome, and schizophrenic hallucinations. In these situations, the stimulus or event may be monitored (Bohlhalter, et al. 2006; Gates, et al. 2004; Neuner, et al. 2007) or reported by the subject (Bar, et al. 2002; Bentaleb, et al. 2002; Lennox, et al. 2000; Oertel, et al. 2007; Shergill, et al. 2001); but this may not always be reliable or practical. In the case of epilepsy, MRI-compatible EEG hardware and analysis software (Gotman, et al. 2004; Hamandi, et al. 2004; Laufs, et al. 2008) are required to detect the timing of the interictal activity at the scalp.
Several model-free analysis methods, such as independent component analysis (ICA), principal component analysis (PCA) and hierarchal clustering analysis (HCA) can be used to differentiate the fMRI signal into individual source components in order to localize specific stimulus responses. These methods have the advantage of not requiring known stimulus timing and have been successfully applied in several fMRI studies (Beckmann and Smith 2004; Calhoun, et al. 2003; Cordes, et al. 2002; Moritz, et al. 2005; Quigley, et al. 2002; Stanberry, et al. 2003; Viviani, et al. 2005). However, these methods inherently result in a large number of components when applied to whole brain fMRI datasets with large numbers of voxels. The classification of these many components and identification of those of practical significance can be a challenge when there are no specific a priori spatial or temporal targets (De Martino, et al. 2007; Rodionov, et al. 2007).
Temporal clustering analysis (TCA) is another data-driven fMRI analysis technique which can be used to determine the timing of stimulus responses (Gao and Yee 2003; Liu, et al. 2000; Lu, et al. 2006; Yee and Gao 2002; Zhao, et al. 2007). The TCA method was first introduced by Liu et al. (Liu, et al. 2000) to find the temporal response and location of eating-related neural activity. In the original TCA methods, the signal of interest was identified as each voxel's maximum signal across time. These signal maxima across the brain were mapped into a histogram where the x-axis indicated the time point and the y-axis showed the number of voxels in the brain that had a maximum at that time point. Peaks in the histogram indicated times at which many voxels reached their maxima simultaneously suggesting timing of stimulus responses. The simplicity of the method prompted our interest in developing it for the localization of interictal epileptiform activity without the use of EEG. Our earlier attempts in adapting this algorithm for temporal lobe epilepsy patients were promising (Morgan, et al. 2004), but revealed a fundamental susceptibility of the technique to artifacts, particularly to head motion (Hamandi, et al. 2005).
To overcome this problem we later introduced the 2dTCA algorithm and evaluated its validity in computer simulations (Morgan, et al. 2008b). This modification of the original algorithm can create separate temporal signals that may be from different sources, thereby separating signals of interest from similar signals caused by motion, noise or other confounds. The result may identify multiple temporal signals much like the multiple ICA components. However, the potential advantage of 2dTCA over ICA in certain applications is that the algorithm will cluster only signal types of interest (i.e. transient spiking), while decreasing sensitivity to signals of other temporal characteristics, thereby reducing the number of components or signals which need to be classified. In two preliminary studies, the 2dTCA method was applied to temporal lobe epilepsy (Morgan, et al. 2007) and to normal controls at rest (Morgan, et al. 2008a) to detect and localize transient spiking with reasonable and promising results. However, before further attempts are made to use this method for localizing interictal activity in epilepsy, more validation of the algorithm for detecting irregular, transient activation in vivo is required.
In this study we applied the 2dTCA technique to detect the timings and spatial locations of sparse, irregular, transient activations of the visual, auditory and motor cortices in normal volunteers. Experiments with one and two independent types of stimuli were employed. Event-related activation using known timing was compared to event-related activation using 2dTCA detected timing in individuals and across groups.
Twelve normal healthy volunteers (8M, 4F, mean age ± stdev: 23.9 ± 7.6 years) participated in this study. All were right handed by self-report. No subjects had a history of neurological, psychiatric or medical conditions as determined by interview. Informed consent was obtained prior to scanning in accordance with Institutional Review Board guidelines.
All imaging was performed on a 3T MRI scanner (Philips Medical Systems, Best, Netherlands) with an 8-channel head coil. Subject motion was minimized using pads and headphones placed within the head coil. All stimuli were programmed using Matlab with the Psychophysics Toolbox (The Mathworks, Inc., Natick, MA). Visual display was provided by MRI-compatible backprojection (Avotec Inc., Stuart, FL) onto a small screen behind the head coil. The headphones within the MRI system provided auditory stimuli.
Structural imaging included a three-dimensional, high-resolution, T1-weighted volume (1mm3) and a two-dimensional, high-resolution, T1-weighted image set in the same slice locations as the functional scanning (0.94 mm × 0.94 mm × 5 mm). Functional BOLD scanning was performed using a T2*-weighted gradient echo, echo planar imaging sequence (matrix = 64×64, FOV = 240 mm, TE = 35 ms, TR = 2 sec, 4.5 mm thickness, 0.5 mm gap, 30 axial slices, EPI reduction factor = 1.8).
Six functional tasks were performed by each subject (Figure 1). Visual block (VB) – Five blocks consisting of 20 seconds of black and white checkerboard flashing at 8 Hz on a gray background contrasted with 20 seconds of fixation on gray background for total of 200 second scan time. Auditory block (AB) – Five blocks of 20 seconds of 2 Hz auditory beeping contrasted with 20 seconds of silence (except for scanner acoustic noise) for a total scan time of 200 seconds. The auditory beeping alternated between two different frequency tones to distinguish it from the scanner noise. Subjects were instructed to tap fingers on both hands at the same frequency as the beeping. Visual spiking 1,2 (VS1, VS2) – Visual fixation on a gray checkerboard with seven (VS1) or six (VS2) transient, visual stimuli presented randomly during the 400 second scan. The transient stimuli were the same flashing checkerboards shown during the VB for a duration of only one second for all stimuli except one which was lengthened to 2 seconds. For both of these scans subjects were instructed to watch the display and look for the longer stimulus. Visual/auditory spiking 1,2 (VAS1, VAS2) – Visual fixation on a gray background with six (VAS1) or seven (VAS2) transient, visual stimuli as in VS1 and VS2, and seven (VAS1) or six (VAS2) auditory two-toned one second beeps as presented in the AB. Subjects were instructed to watch the display and listen for the beeps and to tap twice with both hands each time they heard the auditory beeps. These spiking stimulus timing profiles were the same for all subjects.
The block design functional BOLD scans were acquired to provide robust localization of visual (VB), and auditory and motor (AB) cortices for comparison of the event-related analyses. VS1 and VS2 were acquired to compare the ability of the 2dTCA methods to detect and localize one type of transient activation (visual) during a scan to direct event-related fMRI analysis when the timing is known. VAS1 and VAS2 allowed investigation of the 2dTCA method when more than one type of activation is present at different times and locations.
All functional BOLD images were analyzed using the following procedure using SPM5 software [http://www.fil.ion.ucl.ac.uk/spm/software/spm5/]. First, images were corrected for slice timing differences, motion corrected and spatially normalized to the MNI template (Brett, et al. 2002) using co-registration to the three-dimensional and two-dimensional T1-weighted structural image sets as intermediate steps. Images were then interpolated to 4 × 4 × 4 mm (47 × 56 × 46 voxels).
Each VB and AB dataset was analyzed using the general linear model (GLM) SPM5 using the stimulus timing convolved with the canonical hemodynamic response function as the regressor of interest, and the time courses of the six motion parameters as confounds. The VS1 and VS2 datasets were analyzed together using a fixed-effects GLM with the known transient stimulus timing convolved with the canonical hemodynamic response as the regressors of interest and the six motion parameters of each scan as confounds. Similarly, the VAS1 and VAS2 datasets were analyzed in a fixed-effects GLM using the known visual transient stimuli convolved with the canonical hemodynamic response and the known auditory transient stimuli convolved with the canonical hemodynamic response as the two regressors of interest with the six motion parameters as confounds for each scan. These analyses using known block and event timing yielded five activation maps for each subject: the VB map, the AB map, the event-related visual spiking map from the VS scans (ER-V1), the event-related visual spiking map from the VAS scans (ER-V2), and the event-related auditory (and motor) spiking map from the VAS scans (ER-A2) (Figure 1, bottom).
The individual's VB maps from all subjects were entered into a one-sample t-test to create a group VB activation map. The same was performed for the other four maps to create group maps of AB, ER-V1, ER-V2, and ER-A2.
This analysis was derived from that described in previous work (Morgan, et al. 2007; Morgan, et al. 2008a; Morgan, et al. 2008b) and was carried out with Matlab software [The Mathworks Inc., Natick, MA] using some SPM5 functions. Before the actual clustering was performed, the images were preprocessed to minimize BOLD changes not related to the transient stimuli. The preprocessing involved spatial smoothing with a 8 × 8 × 8 mm full-width half maximum Gaussian kernel, thresholding out non-brain voxels, temporal smoothing of each voxel time course using a running average of every 3 points, linear detrending (Lowe and Russell 1999), and intensity normalization to the first five time points of the voxel to determine percent signal change at each time point. A global time course was calculated as the average time course of all voxels in the brain. This global time course (gtc) and the six motion parameters determined by the SPM5 motion correction process, were then regressed out of the data using linear regression (Macey, et al. 2004). The clustering algorithm was then used to detect transient BOLD spiking in the residual time series data.
Under the assumption that the preprocessed data contains temporal noise that is generally randomly distributed, significant transient spikes in the signal can be defined as times where the BOLD signal fluctuates at least two standard deviations from the mean. To perform the 2dTCA clustering analysis, only those voxels containing possible transient BOLD signal changes were used. These voxels were identified as those which contained at least one time point whose maximum intensity was at least two standard deviations above the mean through time, whose maximum percent signal change is between 0.5 and 8% (signal change may be BOLD related (Handwerker, et al. 2004)), and whose standard deviation through time is less than 1.5. The validity of these criteria was evaluated based on analysis of time courses of activated voxels determined by the event-related analysis using known timing. Figure 2 shows the values of the activated voxels in the visual cortex using the known stimulus timing from one subject (p<0.00001 with cluster size 5) for each of these parameters. Figure 2a shows that 95% of the activated voxels (71% of non-activated voxels) have a maximum percent signal change between 0.5 and 8%. Figure 2b shows that 100% of the activated voxels (97% of non-activated voxels) have a standard deviation through time of less than 1.5. Similarly, 99% of activated voxels (98% non-activated voxels) have a maximum that is at least 2 standard deviations above the mean through time. These parameters may be modified if analyzing data from different field strength MRI scanners or using different pulse sequences. For example, if using a 7T MR scanner the percent signal change window may be increased. If using spin echo imaging at 7T the percent signal change window may be decreased.
In the original TCA methods (Liu, et al. 2000; Lu, et al. 2006), peaks in the histogram indicated times at which many voxels reached their maxima simultaneously suggesting the occurrence of stimulus responses. Using this method, signal intensities due to several different sources would be included in a single histogram. When this histogram is used as a regressor in the GLM, the individual sources of signal changes may not be identified. In the 2dTCA algorithm, we attempted to group together transient spiking profiles that are similar, while keeping dissimilar ones separate. This concept was developed and illustrated in our previous work (see Figure 1 in (Morgan, et al. 2008b)). In this case, the similarity parameter chosen was the time of signal maximum. A two-dimensional mapping was performed for each voxel time course, Z(1:N) (where N = number of time points in one BOLD series), into an N by N matrix, hist2d, by incrementing the values in the following manner for all voxels:
where threshold = mean(Z) + 1.5*standard deviation(Z) and x = the time point at which Z is at its maximum. In this way, each column of the matrix consists of the histogram of significant signal increases for those voxels whose maximum occurred at the same time point.
Similar to the one-dimensional method, each column of hist2d can be considered a histogram with peaks indicating times at which many voxels reached a simultaneous signal increase, thus suggesting a stimulus response. Therefore, each column could be used as a regressor in the GLM. In order to reduce the number of regressors, columns of hist2d which represent significant spiking were determined. We have accomplished this by using peaks in the diagonal of the hist2d matrix to identify columns of interest. For this application, significant peaks in the diagonal of hist2d were defined as those points at least one standard deviation above the mean. Many of these significant columns are very similar to each other, so these columns were then summed when they shared at least 30% of their transient peaks. This reduces the number of histograms and similarity between final histograms. The resulting columns were normalized by subtracting the mean and dividing by the standard deviation. The results were saved as 2dTCA regressors of interest to be used together in a GLM with the motion parameters and gtc included as confounds.
This 2dTCA analysis on each dataset pair (e.g. VS1 and VS2) from each subject produced one or more activation maps. The mean number of activation maps across all twelve subjects was reported. For the VS results, the map that included the visual cortex was identified (2dTCA-V1). The visual cortex activation maps from all subjects were then entered into a one-sample t-test to create the group 2dTCA-V1 map. The same analysis was used to identify both the visual map (2dTCA-V2) and auditory/motor cortex map (2dTCA-A2) from the VAS data (Figure 1, bottom). These were used to create the group 2dTCA-V2 map and the group 2dTCA-A2 map. These allow for qualitative comparison between known timing maps and 2dTCA determined timing maps on the group level.
To quantify the performance of the 2dTCA algorithm in localizing the spiking for each individual, we converted the results from all measures into a common scaling and then used correlation to test similarity of the voxel-by-voxel magnitudes. Specifically, we performed a correlation analysis on the beta weight maps from the GLM analysis. To do this the beta values from the three-dimensional beta images were transformed into one-dimensional arrays, excluding voxels outside the brain. These one-dimensional arrays of beta values were then compared using linear correlation and converted to a Fisher's Z value that is normally distributed with a standard deviation of 1 using the equation: Z=atanh(r)* sqrt(N-3) where r is the Pearson's correlation and N is the number of samples used in the correlation. The correlations between the VB maps and the ER-V1 for each subject were calculated and averaged over the twelve subjects. The correlations between the VB maps and the 2dTCA-V1 map from each person were also determined and averaged in the same way. For comparison, the activation map from a control regressor made up of 20 second blocks of alternating off/on convolved with the hemodynamic response was created and this beta map was also correlated with the VB map for each subject. This block paradigm should have little relationship to the sparse event timing data. This was created as an additional control whose map is not expected to correlate with the VB map. The three sets of correlations across subjects (VB to ER-V1, VB to 2dTCA-V1, and VB to control) were compared using an ANOVA and pairwise Scheffe tests with SPSS software (version 17.0, SPSS, Inc., Chicago, IL) to determine significant differences between the maps.
The same correlation analysis was performed between the VB maps and the ER-V2, 2dTCA-V2, and control maps using the results from the VAS1 and VAS2 datasets. The ANOVA and Scheffe tests were used to determine any significant differences in these correlations across subjects. The auditory/motor cortex maps were compared by performing correlations between the AB maps and the ER-A2, 2dTCA-A2 maps and the control maps in the same way.
The correlation analysis above was also used to quantify the effect of varying parameters of the 2dTCA algorithm. For each test, only a single parameter was changed for analysis of the VS data. The correlation of the map from this modified algorithm to the VB map was calculated over the first six subjects and the average and standard deviation were reported. Several parameters were tested in this way. To investigate the preprocessing, three trials were performed: (1) the spatial smoothing was omitted, (2) the temporal smoothing was omitted, and (3) the temporal smoothing was modified to a five point rather than a three point average. To explore the clustering parameters, the similarity measure was changed from the time of signal maximum to the time of first transient signal change greater than 1.5 standard deviations above the mean (as used in previous versions of the 2dTCA algorithm (Morgan, et al. 2007; Morgan, et al. 2008a; Morgan, et al. 2008b)). Also, the criteria of possible voxels entered into the 2dTCA algorithm were modified from voxels with a maximum change of at least two standard deviations above the mean to one standard deviation and to three standard deviations above the mean. Similarly, the criterion of the standard deviation of the voxels time course being less than 1.5 was omitted. Finally, to study the influence of the identification of significant columns in hist2d, the threshold of the diagonal was changed from one standard deviation above the mean to two standard deviations and 0.5 standard deviations above the mean.
Block and event-related fMRI maps were created for all subjects using the known timing as described above and in Figure 1. The group maps across all subjects using this known timing are shown in Figure 3. The 2dTCA analysis yielded multiple activation maps for each of the VS and VAS pairs of datasets in each subject. The average number of activation maps found using the 2dTCA algorithm on the VS data across all twelve subjects was 3.75 (range 2-5). The four maps from one subject are given in Figure 4. The map from each subject that was qualitatively determined to be the visual cortex map was entered into the one-sample t-test to create the group 2dTCA-V1 map. The average number of activation maps found using the 2dTCA algorithm on the VAS data across all twelve subjects was 5.33 (range 3-8). Four maps from one subject are given in Figure 5. The map from each subject that was qualitatively identified as the visual cortex map was entered in to the one-sample t-test to create the group 2dTCA-V2 map. Similarly, the map identified as the auditory and motor cortex map was used to create the group 2dTCA-A2 map. These group 2dTCA maps are shown in Figure 6.
The quantitative comparison between the event-related analysis using known timing and the 2dTCA detected timing across all subjects is shown using the correlation statistic in Figure 7. In Figure 7a, the correlation between the VB map and the ER-V1 map, VB and 2dTCA-V1 map, and VB and control map are plotted as mean ± standard deviation across all subjects. According to the ANOVA and the pairwise Scheffe tests, there was no statistical difference between the ER-V1 and the 2dTCA-V1 correlated with the VB map. However, these both were significantly more correlated to the VB than the control map (p < 0.001). The same pattern emerged from the VAS results. There was no statistical difference in the correlation between the ER-V2 with VB and 2dTCA-V2 with VB, but both were more correlated to VB than control (p <0.001) (Figure 7b). Finally, Figure 7c shows that the correlation to the AB map was not different across subjects for the ER-A2 and the 2dTCA-A2, but higher than the correlation to the control (p ≤ 0.005).
Figure 8 shows an example of the regressor output from the 2dTCA algorithm for the VS1 dataset from one subject. This histogram is used as a regressor in the GLM analysis. The known times of the transient visual display with the hemodynamic response delay are indicated by the shaded bars. For comparison, the average time course of all activated voxels (p <0.001 uncorrected with cluster size 5) for the same subject is also shown.
The results from the 2dTCA parameter analysis using correlation are given in Figure 9. The bars represent the mean correlation across six subjects and the errors bars indicate the standard deviation. The bars are colored according to general categories of parameters in the algorithm. The 2dTCA algorithm used above and the ER-V1 analyses are given in medium gray bars and are shown for comparison of the results of the modifications. The preprocessing modifications (no spatial smoothing, no temporal smoothing and 5 point temporal smoothing) are shown with diagonal striped bars. The change in similarity measure to first transient signal change greater than 1.5 standard deviations above the mean is given in black. The crossed bars represent the changes in the definition of voxels that are entered into the clustering algorithm from two standard deviations above the mean to three standard deviations above the mean and to one standard deviation above the mean. Also, the criterion of the standard deviation through time of less than 1.5 was omitted from the definition of the transient peak. Finally, the changes in the choice of significant peaks in the diagonal of hist2d are shown in white bars. In 9b, the number of average number of histograms (activation maps) resulting from the corresponding 2dTCA analysis are given. The error bars here represent the range of values across the 12 subjects.
The 2dTCA method is a data-driven, model-free fMRI technique developed to detect and localize transient fMRI activation events. In previous work we validated the ability of the 2dTCA method to detect transient fMRI activation in computer simulated data (Morgan, et al. 2008b). This study represents the next step in the validation process. Our results showed that 2dTCA detected and localized transient activation as well as using known event timing when compared to block-design activation maps in these subjects.
Visual comparison of Figure 6 to Figure 3 reveals the general relationship of the activation maps generated using known timing to those using the 2dTCA timing. The result from the VS data shows that when only the visual stimulus was presented, the group map from all subjects using the known timing includes the occipital activation shown in the block-design group map, as well as some inferior frontal gyrus, thalamus, cerebellum, cingulate and parietal activation. These regions are consistent with those found in response to rare visual targets (Clark, et al. 2000) and may be due to vigilance, arousal, or memory. The 2dTCA group map appears more like the block-design map and does not include these other regions of activation.
The result from the visual activation in the VAS data has a similar pattern, with the event-related group map with known timing having more inferior frontal gyrus, anterior cingulate, cerebellum and thalamus activation. The 2dTCA group map of visual activation from these data includes the same regions as the block-design map, but the activated regions are less extensive. The results of the auditory/motor component of the VAS data are less easily interpreted. Using the known timing, the group event-related activation is similar in location to the auditory/motor block-design map, but much more extensive in each region. Also, the event-related activation map contains strong activation in the anterior cingulate, not in the block-design map. The 2dTCA map very closely resembles the block-design map with focal activation in expected auditory and motor cortices.
The group maps, while informative, do not completely describe how well the 2dTCA algorithm can detect and localize the activation in a single subject. This is important in applying this method to individual epilepsy patients. Therefore, we used the correlation analysis between the event-related maps and the block-design maps for quantification. Figure 7 indicates that across the group, the 2dTCA individual maps correlated with the individual block-design maps as well as the event-related maps using known timing. When compared to the map using the control, which is not expected to resemble the block-design, the correlation is significantly decreased. This holds true in all three cases, but is most significant in Figure 7a. It should be noted that in Figure 7b the pair-wise analysis determined that the increase in correlation of ER-V2 almost reached statistical significance over the correlation of the 2dTCA-V2 with the VB (p=0.07).
Like ICA methods, the 2dTCA algorithm usually results in multiple “components” or activation maps. However, these results show the number of maps from 2dTCA to be relatively small (average 3-5) because they are limited to those with transient activation. Using spatial ICA on comparable size datasets of self-paced finger tapping and resting epilepsy patients, the number of resulting components reported was approximately 100 (Moritz, et al. 2005) and 80 (Rodionov, et al. 2007), respectively. Even though the number of components using the 2dTCA method seems to be significantly reduced, the selection of the component(s) of interest remains an issue. In Figure 4, the four activation maps from a single subject included the expected visual cortex map as well as a map that appears to correspond to the “default-mode” network (Greicius, et al. 2003; Raichle, et al. 2001). The other two maps resemble previously defined resting-state networks found by ICA (Damoiseaux, et al. 2006; De Luca, et al. 2006) and may be due to modulation in attention, internal dialog or working memory throughout the acquisition. In Figure 5, the expected visual and auditory/motor activation maps were found along with two other maps. These other two maps appear less likely to be activation of known resting-state networks than motion or physiological noise. Qualitative assessment across all subjects showed many of the same activation maps as those shown in these two subjects.
While the accuracy of temporal profile of the events detected by 2dTCA is implied by the similarities of the maps, Figure 8 gives an example of a regressor determined by 2dTCA. The greatest peaks in the regressor are concurrent with the known times of the events. However, there are other peaks in the 2dTCA regressor which may reflect true variability of the BOLD signal in the activated region (see Figure 8 bottom) such as due to attention or arousal. These signal changes may also be artifacts resulting from motion, physiological noise (Bhattacharyya and Lowe 2004; Birn, et al. 2006) or other activity in the same region. Also, the removal of the global time course prior to implementing the 2dTCA algorithm may introduce anti-correlations across regions (Murphy, et al. 2008). This effect is expected to be minimal in this analysis, but is currently unverified.
The correlation results in Figure 9 show that the 2dTCA algorithm seems to be most sensitive to the similarity measure chosen to group histograms (dark bars). When the similarity measure was changed, the algorithm had fewer numbers of resulting activation maps to consider, but correlations with VB varied widely across subjects. The algorithm is also sensitive to temporal smoothing (diagonal striped bars) and threshold chosen for significance in the diagonal of hist2d (white bars) which result in a general drop in correlation to VB with no change in number of histograms. The algorithm seems to be less sensitive to changing the number of standard deviations above the mean considered as a transient peak to one instead of two (crossed bar). This suggests that regardless of whether you define transient peaks as one or two standard deviations above the mean, these peaks will cluster in the histogram with increases in the diagonal of hist2d at times of transient activation resulting from our stimulus; however, the parameter chosen to identify these increases in the diagonal will determine the accuracy and sensitivity of the algorithm.
In conclusion, these results suggest that the 2dTCA algorithm can accurately detect and localize sparse, irregular, transient events of unknown timing in individual subjects. In all subjects, the algorithm detected the expected activation even when more than one type of stimulus was presented. The specificity of the 2dTCA algorithm to transient events may reduce the number of detected components from those found using other standard data-driven techniques, potentially making this a more practical method for this detection of sparse transient activation using fMRI. Finally, this method is not recommended to be used instead of known timing, but may provide complementary information to such analyses.
This work was supported in part by NIH EB00046 and NIH R01 1NS055822.