Surface-based analysis increases the specificity of cortical activation patterns and connectivity results

Abstract

Spatial smoothing of functional magnetic resonance imaging (fMRI) data can be performed on volumetric images and on the extracted surface of the brain. Smoothing on the unfolded cortex should theoretically improve the ability to separate signals between brain areas that are near together in the folded cortex but are more distant in the unfolded cortex. However, surface-based method approaches (SBA) are currently not utilized as standard procedure in the preprocessing of neuroimaging data. Recent improvements in the quality of cortical surface modeling and improvements in its usability nevertheless advocate this method. In the current study, we evaluated the benefits of an up-to-date surface-based smoothing in comparison to volume-based smoothing. We focused on the effect of signal contamination between different functional systems using the primary motor and primary somatosensory cortex as an example. We were particularly interested in how this signal contamination influences the results of activity and connectivity analyses for these brain regions. We addressed this question by performing fMRI on 19 subjects during a tactile stimulation paradigm and by using simulated BOLD responses. We demonstrated that volume-based smoothing causes contamination of the primary motor cortex by somatosensory cortical responses, leading to false positive motor activation. These false positive motor activations were not found by using surface-based smoothing for reasonable kernel sizes. Accordingly, volume-based smoothing caused an exaggeration of connectivity estimates between these regions. In conclusion, this study showed that surface-based smoothing decreases signal contamination considerably between neighboring functional brain regions and improves the validity of activity and connectivity results.

Figure 1

Individual cortical activation pattern (SPM maps) due to tactile stimulation and illustration of the transformation of a 3D volume to surface data. (Upper row) Individual cortical activation pattern (SPM maps) induced by tactile stimulation of the fingers of the right hand using a 9 mm FWHM (full width at half maximum) Gaussian smoothing kernel. The arrows point to 2 exemplary locations within the primary motor and the primary somatosensory cortex. However, being located in 2 different functional (motor vs. sensor) and anatomical (pre- vs. postcentral gyrus) structures, their Euclidean distance within the 3D space was only approximately 4 mm. (Lower) Illustration of the transformation of a 3D volume to surface data. The 2 exemplary locations within the pre- and postcentral gyrus in the original 3D volume are approximately 4 mm apart. By mapping the 3D volume to the surface space (FreeSurfer spherical template), the same 2 points were approximately 40 mm apart.

Figure 2

Schematic of different GLM analysis approaches. All approaches started from the same fMRI data. Consequently, a slice time correction, a realignment and a co-registration to the corresponding structural T1 image were performed. (1) In the first approach, the functional images were normalized and smoothed (to 6, 9 and 12 mm) before applying the GLM. (2) In the second approach, the GLM was performed before normalization and smoothing. (3) In the third approach, the GLM was performed in the individual space (as in approach 2). Afterwards, the results were mapped to the individual surface and normalized to a standard surface. The smoothing occurred on the normalized surface. Group analyses for all 3 approaches were performed using a 1 sample t-test.

Figure 3

Schematic of HRF simulation. rs-fMRI data were co-registered to their corresponding structural t1 image. Using a DARTEL template in normalized space, a region of interest (ROI) was defined within the primary somatosensory cortex (S1); values of the ROI ranged from 1 (center) to 0 (6 mm distance) to modify the amplitude of the simulated HRF. For each subject an individual HRF signal was created using the simTB-Toolbox. After mapping the ROI to the individual space, the simulated HRF signals were added to individual rs-fMRI time series. Afterwards, the GLM analysis was performed within the individual space; consequently, the volume (2) and surface-based (3) group analyses were carried out.

Figure 4

Functional connectivity analysis of simulated fMRI signals within the pre- and postcentral gyri. Upper part: Schematic of the creation of a simulated 0.05 Hz signal (signal 1) in the postcentral region. Correlated (2c) and uncorrelated (2 u) signals placed into the precentral region. Lower part: Results of the functional connectivity analysis between signal 1 and the correlated (2c) as well as the uncorrelated (2 u) signal using the VBA and SBA. The difference between the designed correlation (r = 0.82 for 1~2c and r = 0.04 for 1~2 u) and the actual measured correlation is shown in the lower row. Significant differences between the VBA and SBA are indicated.

Figure 5

Comparing cortical activation patterns in specific brain regions of volume and SBAs induced by tactile stimulation of the right hand. Cortical activations were counted in 4 brain regions (rows): precentral gyrus, central sulcus, postcentral gyrus and sulcus and results for the 3 different approaches (aligned in columns: 1–2 volume-based, 3 surface-based) were compared. Again, the results of the left cortex are shown (corrected for multiple comparisons using TFCE and adjusted at p ≤ 0.01 FWE).

Figure 6

Results of the SPM analysis of a simulated HRF signal. The simulated BOLD signal occurred every 10 s and lasted 1 s. GLM-results were smoothed using 6, 9 and 12 mm; 2nd level results were corrected for multiple comparisons and adjusted at p ≤ 0.05 FWE. The number of active voxels within the precentral and postcentral gyri are displayed for each separate analysis.