There is great interest in the study of brain connectivity (both structural and functional), and on the development of methods that facilitate these investigations [1]. In functional connectivity (FC), there is also growing interest in characterising the dynamic changes (i.e. so-called time-resolved or dynamic FC, dFC), with several studies showing dynamic patterns of fluctuation in time-dependent FC measures [2]. The structural and functional aspects of connectivity are most often considered independently and only linked at a late stage of the analysis. In contrast, track-weighted FC (TWFC) was proposed as a means to combine the structural and (static) functional information into a single image, by integrating a given functional network with a diffusion MRI tractogram [3]. Here we propose TW dynamic FC (TWdFC), by extending TWFC in two ways: first, it does not rely on defining an a-priori FC network; second, it allows studying dFC.

In TWdFC, each streamline in a tractogram is assigned the functional correlation between the fMRI data at its end-points (Note: this approach is facilitated here by the well-defined track’s end-points in the grey/white matter (GM/WM) interface using the anatomically-constrained tractography (ACT) framework [4]). Importantly, TWdFC does not consider correlations over the entire fMRI time-series, but rather over a sliding-window [5]; Pearson cross-correlation was used for the results shown here. As with standard TWFC, FC values associated with all tracks traversing a given voxel are averaged to produce the final intensity (Note: this voxel can be smaller than the acquired voxel, i.e. with super-resolution [6]). A (3D+time) TWdFC data-set is thus generated. For comparison, the same approach but without sliding-window was also used (referred to as TW static FC, TWsFC).

Data acquisition: Eight healthy volunteers were scanned on a 3T Siemens Trio scanner. BOLD fMRI was acquired using GE-EPI (TE/TR=30/3000ms, voxel-size=3mm isotropic, 200 volumes). Diffusion MRI was acquired using twice-refocused SE-EPI (60 directions, b=3000s/mm², voxel-size=2.5mm isotropic). Reference EPI data with opposite phase-encoding were acquired for both fMRI and diffusion MRI to correct for susceptibility distortions [3], and anatomical T1-weighted images (0.9mm isotropic) were acquired for tissue segmentation [4].

BOLD processing: The (distortion corrected) BOLD data were realigned to the (corrected) b0 diffusion MRI image, followed by motion correction, smoothing (FWHM=8mm), detrending, band-pass filtering (0.01-0.1Hz), and nuisance regression (motion, WM and CSF).

Diffusion processing (using MRtrix [7]): Fibre orientation distributions (FODs) were estimated using CSD [8] (lmax=8). T1 data were used for segmentation (using FSL [9]), and co-registered to diffusion data. Whole-brain fibre-tracking (10 million) was done using probabilistic streamlines and the ACT framework [4].

TWdFC was calculated with 1min sliding-window [10] and 0.9mm super-resolution voxels. The whole-brain-average root-mean-squared (RMS) difference between TWdFC and TWsFC was calculated as a summary measure of the dynamic changes. Dynamic changes were also summarised by calculating the voxel-wise 95% confidence interval of the temporal variations (TWdFC_95%CI). To illustrate spatial dynamic features in the data, the individual’s TWdFC data were analysed using independent component analysis (ICA) (using FSL).

Figure 1 shows the summary RMS difference for each subject, demonstrating the dynamic nature of the data. Figure 2 shows example time-frames from an individual’s TWdFC map, and the corresponding TWdFC_95%CI (as well as TWsFC for reference, and directionally-encoded track-density image (DEC-TDI [6]) for anatomical reference). Different WM pathways have very different temporal characteristics, corresponding to fluctuations in the GM regions they link (see Figure 3 for example time-courses for four manually-drawn WM regions); e.g. fluctuations in optic radiations reflect the dynamic FC correlations between the lateral geniculate nucleus and the primary visual cortex. Figure 4 shows illustrative ICA spatial components (with IC numbers based on ranking of explained variance) for the same subject, and their associated time-courses. Figure 5 shows a ‘parcellation’ of the corpus callosum based on the ICA decomposition.We described a methodology for super-resolution TWdFC mapping, which fuses structural and functional connectivity data into a quantitative 4D image, providing a new means to investigate dynamic connectivity. In contrast to most FC-based methods (which focus on GM), TWdFC operates in WM, and its intensity reflects the (dynamic) correlation of the GM regions that the corresponding WM pathways connect. The structural connectivity information effectively ‘constrains’ the extremely large number of possible connections in the FC data (i.e. each voxel’s connection to each other), thus providing a way of reducing the problem’s dimensionality, while still maintaining key features [2]. In its current form, only contributions from direct connections (i.e. with a direct structural link) are considered, although the methodology can in principle be extended to incorporate also higher (e.g. second order) connections [11].