Figure 1.
General conceptual difference between the non-parametric ISC and parametric GLM analysis.
The ISCs are computed voxel-wise over the measured time series of every possible subject pair and then the results are combined to a single statistic. The GLM analysis fits the mathematical model (Here: boxcar function convolved with the canonical hemodynamic response function (HRF)) to the measured time-series of every subject and the group level results are then combined from the results of the individual subjects' analyses.
Table 1.
FDR corrected GLM thresholds for different tasks.
Table 2.
Voxel-wise correlation measures, Eq. (3).
Figure 2.
The correlation measure and the Dice index.
The bars show the correlation measure between ISC and GLM and the lines present the Dice index values from different significance levels. The continuous black line presents the average over the Dice values within the current task. The HA task has higher correlation measure than other tasks and a high Dice index value. The EO task has the lowest correlation measure and the Dice index is also lower and varies the most with the thresholds. This suggests that a high correlation measure predicts a high Dice index value. We note that the values used as the basis for this figure are of higher numerical precision than those reported in Tables 2 and 3.
Table 3.
Dice Indices, Eq. (4).
Figure 3.
GLM and ISC analysis results for the AN task (thresholded and FDR corrected, q
0.05 (a), q
0.005 (b), q
0.001 (c) ). In the images, the red color indicates voxels which are activated according to both ISC and GLM methods, blue indicates voxels activated according to GLM but not according to ISC and green indicates voxels activated according to ISC but not with GLM. The images are in neurological orientation. There is a notable correspondence between the ISC and GLM maps especially in auditory cortex, visual cortex, and cingulate gyrus. We can also see that the ISC analysis was clearly more liberal than the GLM analysis with a loose threshold (q
0.05), but became more conservative when the thresholds became tighter (q
0.005 and q
0.001).
Figure 4.
GLM and ISC analysis results for the AN task.
The scatterplot (a) presents the voxel-wise statistic values of GLM (horisontal axis) and ISC (vertical axis). Red lines define the thresholds with levels q = 0.05, q = 0.005 and q = 0.001. The second image (b) displays the corresponding histogram, which shows more clearly how the mass of the values is distributed with respect to the thresholds defined by the red lines. Most of the values are focused close to the origin which is not visible in the scatterplot.
Figure 5.
GLM and ISC analysis results for the EO task.
In the image the thresholded (FDR corrected, q0.001) results for EO task are presented as a binary overlay image. The color coding in the image is the same as in Figure 3. The threshold images from the levels q
0.05 and q
0.005 are visible in the Figure S1 of the Supplement. Both methods find the same activation areas widely across the brain, including lateral occipital cortex, inferior frontal gyrus, precentral gyrus and supplementary motor cortex. Note also how ISC only (green) and commonly detected areas (red) are vanishing faster than GLM only areas (blue) when the threshold becomes more conservative. Thus, the ISC analysis was more conservative of the two methods especially with the lowest q-value. This tendency explains relatively high variation in the Dice index values with different significance levels for this particular task.
Figure 6.
GLM and ISC analysis results for the HA task.
In the image the thresholded (FDR corrected, q0.001) results for HA task are presented as a binary overlay image. The threshold images from the levels q
0.05 and q
0.005 are visible in the Figure S3 of the Supplement. The color coding in the image is the same as in Figure 3. Here it is clear that commonly detected areas (red) are dominant. There are also a notable number of ISC only detections (green), which might indicate that ISC can detect activations which are not detectable by GLM. On the other hand, some GLM only activations were located in cerebrospinal fluid, which suggested that there might exist measurement artifacts.
Figure 7.
The voxels consistently detected as activated by one method and not by the other with AN task.
Green color indicates voxels which were detected as activated by GLM in all thresholding levels, but not detected as activated by ISC in even the most liberal thresholding level (q = 0.05). Viceversa, blue color indicates voxels which were detected as activated by ISC in all of the thresholding levels, but not detected as activated by GLM with even the most liberal thresholding level (q = 0.05). Mostly these are isolated voxels or voxels lying near the boundary of the activation area. However, the ISC detected activations in Posterior and Anterior cingulate cortex and Precuneus as well as Occipital lobe that were not detected by the GLM. These areas are suspected to overlap with the default mode network in several studies, e.g., [34]–[36].
Figure 8.
Similarity of the detected activation region and ground truth activation region in the simulation study.
The lines present the Dice index values between the simulated versus detected activation area by ISC with different thresholding levels (blue lines) and by GLM with different thresholding levels (red lines). The ISC performed well with lower noise levels (SNR 1/100 and 1/200) but failed with the highest noise level (SNR 1/1000). The GLM performed overall well, but has a lower detection rate at low noise levels compared to ISC. This is due to false positive detections on the areas nearby ground-truth activation areas due to the effects of the spatial smoothing.