Different MRI structural processing methods do not impact functional connectivity computation

Abstract

Resting-state functional magnetic resonance imaging (rs-fMRI) has become an increasingly popular technique. This technique can assess several features of brain connectivity, such as inter-regional temporal correlation (functional connectivity), from which graph measures of network organization can be derived. However, these measures are prone to a certain degree of variability depending on the analytical steps during preprocessing. Many studies have investigated the effect of different preprocessing steps on functional connectivity measures; however, no study investigated whether different structural reconstructions lead to different functional connectivity metrics. Here, we evaluated the impact of different structural segmentation strategies on functional connectivity outcomes. To this aim, we compared different metrics computed after two different registration strategies. The first strategy used structural information from the 3D T1-weighted image (unimodal), while the second strategy implemented a multimodal approach, where an additional registration step used the information from the T2-weighted image. The impact of these different approaches was evaluated on a sample of 58 healthy adults. As expected, different approaches led to significant differences in structural measures (i.e., cortical thickness, volume, and gyrification index), with the maximum impact on the insula cortex. However, these differences were only slightly translated to functional metrics. We reported no differences in graph measures and seed-based functional connectivity maps, but slight differences in the insula when we compared the mean functional strength for each parcel. Overall, these results suggested that functional metrics are only slightly different when using a unimodal compared to a multimodal approach, while the structural output can be significantly affected.

Figure 1

Workflow of the analysis. Each participant underwent resting-state functional, T1-weighted (T1w), and T2-weighted (T2w) MRI scans. Surface reconstruction was performed through a unimodal pipeline (UP; using T1w signal) or a multimodal pipeline (MP; combining T1w and T2w signals). For the UP workflow (pink triangle), 3D T1w structural images were processed using the recon-all processing stream, which performed all reconstruction steps, including motion correction, intensity normalization, skull-stripping, white matter (WM) segmentation, spherical morph, and parcellation. MP workflow (blue and pink triangle) included the T2w images (blue triangle) in the Freesurfer workflow, recomputing spherical morph and cortical parcellation to adjust the pial surfaces. The preprocessing steps for resting-state data included head movement correction, intensity normalization, anatomical registration, and smoothing. For the anatomical registration step, the rs-fMRI preprocessing pipeline was run independently for the UP and the MP. Structural (volume, cortical thickness, and gyrification index) and functional (mean FC, seed-ROI maps, spatial topology, and graph analysis) outcomes were compared between the two different pipelines.

Figure 2

Structural differences between unimodal and multimodal pipelines. Top panel: difference in cortical thickness at the vertex level, stronger within the bilateral insula (first row); differences in the gyrification index at vertex level (second row). Middle panel: significant differences in cortical thickness (left) and gyrification index (right) within frontal, parietal, temporal, and occipital cortices were found between the two processing modalities. Significantly thicker cortical thickness was reported for the multimodal pipeline, compared to the unimodal pipelines. Significantly higher gyrification index was reported for the unimodal pipeline, compared to the multimodal pipelines. Bottom panel: subdivision of the Desikan–Killiany atlas parcels for the four lobes.

Figure 3

Functional connectivity differences between unimodal and multimodal streams. Left panel: differences in cortical thickness, volume, mean functional connectivity (FC) with different smoothing parameters (Gaussian smoothing kernels: 5 mm, 4 mm, and 6 mm), median FC, and mean FC computed using a subcohort of participants acquired with the same MRI protocol. Regions were reordered according to t-values from the paired sample t-test cortical thickness analysis (left hemisphere). The left insula was the parcel with the highest t value (absolute value); the left anterior cingulate (ACC) was the parcel with the lowest t value (absolute value). Right panel: differences in mean FC between unimodal and multimodal streams projected to the Schaefer 100-parcel atlas. A significant difference was reported only for the left prefrontal cortex. T: threshold, S: Gaussian smoothing kernels.

Figure 4

Spatial topological differences between unimodal and multimodal streams. The left insula and left anterior cingulate cortex (ACC) were selected as regions of interest (ROIs) and the Euclidean distance in the 3D anatomical space between ROI’s center of gravity and peak FC voxels was computed. Top panel: Euclidean distance did not report significant differences between unimodal and multimodal streams. Bottom panel: Significant differences were observed between the delta shift (Euclidean distance) of the left insula and the left ACC.

Figure 5

Difference in FC topography between unimodal and multimodal streams. Top panel: mean seed-based connectivity maps for the left insula (left) and left anterior cingulate cortex (ACC) (right). Middle panel: Significant differences between the insula and ACC error maps (computed as the difference between UP and MP). Higher values mean higher errors. Bottom panel: Significant differences between the insula and ACC error maps were found at the voxel level.

Figure 6

Graph theory metrics differences between unimodal and multimodal streams. Betweenness centrality, closeness centrality, clustering coefficient, degree centrality, eigenvector centrality, and normalized strength were computed for each parcel and each subject. Outcomes were organized into a j × k matrix (j = the number of parcels; k = the number of participants) and fed to a principal component analysis. The first component (PC1) was selected and compared between unimodal and multimodal streams. PC1 from UP and MP showed the same structure for the matrix (r > 0.98).