Identification of functional white matter networks in BOLD fMRI

Abstract

White matter signals in resting state blood oxygen level dependent functional magnetic resonance (BOLD-fMRI) have been largely discounted, yet there is growing evidence that these signals are indicative of brain activity. Understanding how these white matter signals capture function can provide insight into brain physiology. Moreover, functional signals could potentially be used as early markers for neurological changes, such as in Alzheimer’s Disease. To investigate white matter brain networks, we leveraged the OASIS-3 dataset to extract white matter signals from resting state BOLD-FMRI data on 711 subjects. The imaging was longitudinal with a total of 2,026 images. Hierarchical clustering was performed to investigate clusters of voxel-level correlations on the timeseries data. The stability of clusters was measured with the average Dice coefficients on two different cross fold validations. The first validated the stability between scans, and the second validated the stability between populations. Functional clusters at hierarchical levels 4, 9, 13, 18, and 24 had local maximum stability, suggesting better clustered white matter. In comparison with JHU-DTI-SS Type-I Atlas defined regions, clusters at lower hierarchical levels identified well-defined anatomical lobes. At higher hierarchical levels, functional clusters mapped motor and memory functional regions, identifying 50.00%, 20.00%, 27.27%, and 35.14% of the frontal, occipital, parietal, and temporal lobe regions respectively.

1. INTRODUCTION

Blood oxygen level dependent functional magnetic resonance imaging (BOLD-fMRI) is a common noninvasive technique used to measure neuronal activity through the spontaneous hemodynamic responses in blood¹. At resting state, BOLD-fMRI is used to examine the baseline functional signaling within the brain during which the subject is not completing any task. Using these signals, grey matter (GM) has been used extensively to create functional maps in the brain². This GM functional mapping is important for understanding aging effects and diseases, such as Alzheimer’s Disease³. In contrast to GM, white matter (WM) is less voluminous and has smaller BOLD signals especially at resting state⁴. As a result, WM BOLD signals have previously been discounted or considered noise, but there is growing evidence these WM signals can provide insight into functional brain regions⁴.

Recent literature has begun investigating the significance of WM signals in the resting state BOLD-fMRI. The hemodynamic responses of WM and GM have been found to be similar in fMRI, providing evidence that WM signals should not be dicounted⁴. Similarly, a recent study has revealed clustered WM correlations are highly consistent with identified GM correlations⁵. Furthermore, WM functional networks have been correlated with GM functional networks to investigate the inter-network connectivity⁶. However, the previous work has been completed using k-means clustering techniques only. Our work uses a hierarchical clustering approach to investigate clustering on BOLD-fMRI WM correlations.

2. METHODOLOGY

From the OASIS-3 database⁷, 711 (cognitively normal, approximately equal numbers of males and females, ages ranged between 42 and 95 years) individuals were selected. Many individuals provide longitudinal data, and most were scanned twice in a session, resulting in a total of 2,026 resting state BOLD-fMRI images.

Slice timing and head motion were removed from the fMRI volumes, and the mean cerebrospinal fluid (CSF) signal and 24 motion-related parameters were modeled as covariates and regressed out from the BOLD signals. The data were then detrended and passed through a temporal filter with a passband frequency of between 0.01 and 0.1 Hz. All these procedures were carried out using a customized pipeline based on the DPABI toolbox^8,9. The Computational Anatomy Toolbox (CAT12) was then used to segment GM, WM, and CSF tissue based on the T1-weighted images¹⁰. Using co-registration and normalizing functions in SPM12¹¹ the filtered fMRI data and tissue masks, were spatially normalized into MNI space (voxel size = 3 × 3 × 3 mm³). A group-wise WM mask was constructed by averaging the WM parcellations (probability maps) that were derived from CAT12 across all subjects and applying a threshold. The initial threshold was set to 0.95. However, this cropped out many important WM voxels, particularly small structures spatially located between GM regions, e.g., internal, and external capsules. We then spatially expanded the WM mask by decreasing the threshold gradually in steps of 0.01 until the overlap between the mask and GM area could be visually noticed on the group mean T1 image. We found that 0.8 was the minimal value that could produce a clean WM mask while retaining most of the important WM structures. In a similar manner, a GM mask was reconstructed using a lower threshold of 0.65 that was specifically selected by setting a threshold of 0.5 and gradually increasing in steps of 0.01 until there were no overlapping voxels between WM and GM. Then, the fMRI data within the WM mask were spatially smoothed with a 4-mm full width at half maximum Gaussian kernel. The preprocessed results were subjected to a manual quality control procedure in which the passing criteria included: (1) all the preprocessed results must be successfully generated; (2) the maximal translations and rotations of head motion must be less than 2 mm and 2°, respectively; (3) the mean frame-wise displacement (FD) must be less than 0.5 mm¹² and (4) the spatial normalization was acceptable by an expert’s visual inspection.

Images were acquired every second over a 154 second interval. Each image and WM mask were reshaped into 2D (spatial coordinates × timestamps). The WM mask was used to identify WM voxels in each image. A voxel level Pearson correlation coefficient was calculated across all WM voxels in each individual image. Then, each voxel level correlation coefficient was averaged across all images to create an average voxel level correlation. Clusters on the WM voxel correlations were computed using MATLAB’s clusterdata and linkage hierarchical clustering functions¹³. The distance between each correlated voxel was measured using one minus the correlation. Then, the computed distances were clustered by minimizing the unweighted average between each distance.

To measure cluster stability, two cross fold validations were completed. The first measured the stability between each scan session to measure the cluster repeatability. One scan from each subject and session was randomly assigned to a fold, resulting in two folds with 1,013 image scans each. The second validation measured the stability between subject populations to measure cluster reproducibility. Subjects were randomly assigned to a fold, and only the first session scans were used. Most subjects have two images in the first session, resulting in two folds with 355 and 356 subjects, and 708 image scans in each. Clustering of correlations between BOLD-fMRI signals at hierarchical levels between 2 and 40 was performed on each fold. The similarity between each individual cluster was measured using the Dice coefficient. The stability was measured as the mean Dice coefficients of all clusters at level.

This clustering technique was also compared to regions identified by the JHU-DTI-SS Type-I Atlas “Eve Atlas”¹⁴. Since our clustering results appeared to be visually symmetric between the left and right hemispheres, we compared our identified regions to the left hemispheric regions in the Eve Atlas. In this comparison, WM was clustered on hierarchical levels of 25 through 65. Each functional clustered region was compared to every Eve Atlas region using the Dice coefficient. The regions in the Eve Atlas and clustered WM with the largest Dice coefficient were considered the best pair and removed from further possible pairings. This process was repeated until all regions were paired. The best Dice coefficients for each region were doubled to create an estimate for individual cluster similarities in both hemispheres. Similarity was measured as the mean of the best Dice coefficients at a level.

3. RESULTS

Clustering on the Pearson correlation coefficients across all WM voxels successfully identified visually clear regions. On the first cross fold validation examining the scan-rescan stability, the average Dice coefficient at every hierarchical level was equal to one. On the second cross fold validation examining the population stability, the average Dice coefficient initially decreased, then becomes relatively stable, as the hierarchical level increased. Dice coefficient local maximums were identified at hierarchical levels 4, 9, 13, 18, and 24 indicating possible better defined functional clusters.

Visualizing the functional clusters identified from the stability plot showed distinct WM clusters consistent with the well-defined anatomical regions of the brain. Clustering appears symmetric across the left and right hemispheres. At all hierarchical levels, the number of clusters are equal to the hierarchical level. At lower hierarchical levels, the major lobes-occipital, parietal, coronal, and temporal- are defined. As the hierarchical level increases, motor, and sensory regions, such as the limbic lobe and motor cortex are defined (Figure 3). It is likely for activation to be identified in these functional regions at resting state.

Figure 3.

Local maximums occur at hierarchical levels of 4, 9, 13, 18 and 24, suggesting distinct brain regions are better identified at these levels (left). When visualizing, at a hierarchical level of 4 the major lobes are defined. As the level increases, motor and sensory regions of the brain are identified, which are expected to be active even at resting state (right).

When the functional clustered WM regions are compared to the Eve Atlas, portions of the frontal, temporal, and orbital gyrus are consistently identified at all hierarchical levels starting at level 4. Hierarchical level 30 is most like the Eve Atlas with the best mean Dice coefficient equal to 0.2881. At this level, 17 of 34 frontal, 2 of 10 occipital, 3 of 11 parietal, and 5 of 14 temporal regions are identified. Portions of the cerebellum and midbrain are also identified. The primary functions of the frontal and temporal lobe are motor, memory, language, and sensory processing, which are consistent with the expected active brain regions at resting state. Within the frontal lobe, 46.05% of the corpus callosum, 37.73% of the orbital gyrus, and 41.66% of the internal capsule regions are captured by the functional clusters and all contribute to functional memory. Even at the best matched Eve Atlas regions, the Dice coefficient is much lower than the Dice coefficients determined from the stability measurement, suggesting this functional clustering is distinct from a structural parcellation (Figure 4).

Figure 4.

Functional clusters at hierarchical level of 30 best capture the regions in the Eve Atlas. At this level, 17 of 34 frontal, 2 of 10 occipital, 3 of 11 parietal, and 5 of 14 temporal lobe regions are identified in the functional clusters. Functional clusters at a hierarchical level of 30 appear highly symmetrical across hemispheres. The functional parcellation is different from the structural parcellation identified by the Eve Atlas (right).

4. DISCUSSION

Our study revealed that WM BOLD signal is composed of functional networks, in much the same way as the more commonly analyzed gray matter BOLD signal. While there is no singular value for ‘the number of WM networks’, we find relative stability of 6–40 clustered regions with reproducible volumetric areas in two-fold cross validation cohort. At all levels, these networks are bilaterally symmetric, with coarse levels generally representing WM within brain lobules, and individual gyri and functionally relevant areas shown at more fine scale. Moreover, these networks do not necessarily align perfectly with well-known WM pathways.

Our findings replicate previous studies which show WM networks that correspond to general patterns of WM regions and pathways^5,15,16. However, our results suggest that these clusters, at all spatial scales, are often distinct from WM regions. For example, average overlap coefficients between the Eve Atlas and functional clusters are generally low to moderate. Thus, there is unique information obtained from WM functional data this is not solely derived or driven by structural connections. Future work should investigate the reliability of these networks under different image acquisition conditions, different clustering methods⁵, and different tasks.

Figure 1.

In resting state BOLD fMRI signals there are expected correlations between GM voxels which can be used to identify anatomical regions⁴ (left). In a similar manner, we hypothesize correlations can be identified between WM voxels (right), providing evidence there may be informative WM correlations in resting state BOLD fMRI data.

Figure 2.

The approach to discover spatial patterns in fMRI imaging using standardized clustering techniques, focusing on WM.