Pipeline Overview
We refer to our pipeline as TheVirtualBrain-UK Biobank (or TVB-UKBB) pipeline. It is built from a fork of the UK Biobank pipeline (Alfaro-Almagro et al., 2018). The UK Biobank pipeline processes a variety of MRI modalities but, for the purposes of creating TVB inputs, we focused on modifying and extending the existing structural (T1w, T2 FLAIR), functional (resting-state, task), and diffusion-weighted MRI sub-pipelines. The processing of other MRI modalities (e.g., susceptibility-weighted imaging) in the TVB-UKBB pipeline remain unaltered and untested.
Figure 1 shows the general workflow of the whole pipeline, its sub-pipelines, and their outputs. The major output of the structural MRI pipeline is the user-defined parcellation registered to the subject’s T1w image. The registered parcellation is used by both the functional and diffusion MRI sub-pipelines to define ROIs for computing average regional timeseries and connectivity measures for TVB inputs. Following the completion of the functional and diffusion MRI sub-pipelines, an ‘IDP’ pipeline computes image-based metrics for all modalities. Finally, our newly developed QC pipeline generates a comprehensive HTML-based report for manual quality assurance procedures.
The pipeline currently supports Cam-CAN and ADNI3 datasets, and can be customized by naive users to support novel datasets. Here we demonstrate usage of the TVB-UKBB pipeline using the Cam-CAN dataset (cite), which includes T1w, T2*, resting-state and task-fMRI, field maps, and dMRI from ~650 adults aged 18-99. In these examples, we used a Schaefer-Tian parcellation consisting of 400 cortical and 20 subcortical regions.
The key TVB inputs generated by the pipeline can be visualized and analyzed with ease. Figure 6 shows the pipeline outputs of interest for connectome-based modelling for an example subject. These include the structural connectivity weights and tract lengths matrices, and the resting-state BOLD-fMRI responses and functional connectivity matrix.
Figure 6. An example subject’s set of pipeline outputs for connectome-based modelling. These include (A) a weights matrix and (B) a tract lengths matrix from dMRI processing that capture the subject’s structural connectivity; (C) a functional connectivity matrix of Pearon correlation coefficients, and (D) the region of interest (ROI) time series from resting-state fMRI processing. The structural connectivity matrices are presented on a log scale to enhance readability. Ten ROIs were chosen randomly for presentation in (D).
Our pipeline generates inputs for connectome-based modelling, with file formats that are directly compatible with TheVirtualBrain (TVB; thevirtualbrain.org) (Supplementary Figure 1). These include the structural connectivity weights and tract lengths matrices, as well as the ROI time series and functional connectivity matrix from resting-state fMRI scans. ROI location information such as hemisphere or subcortical localization and centroid coordinates are also included. Towards the end of the pipeline, these TVB-input files are given the appropriate file names, placed in the correct folder structure, and compressed into a zip file that can be accepted by TVB without further processing. This zip file can be found in the top-level directory for each processed subject.