Converging on consistent functional connectomics

Abstract Functional interactions between brain regions can be viewed as a network, empowering neuroscientists to leverage network science to investigate distributed brain function. However, obtaining a brain network from functional neuroimaging data involves multiple steps of data manipulation, which can drastically affect the organisation and validity of the estimated brain network and its properties. Here, we provide a systematic evaluation of 576 unique data-processing pipelines for functional connectomics from resting-state functional MRI, obtained from all possible recombinations of popular choices for brain atlas type and size, connectivity definition and selection, and global signal regression. We use the portrait divergence, an information-theoretic measure of differences in network topology across scales, to quantify the influence of analytic choices on the overall organisation of the derived functional connectome. We evaluate each pipeline across an entire battery of criteria, seeking pipelines that (i) minimise spurious test-retest discrepancies of network topology, while simultaneously (ii) mitigating motion confounds, and being sensitive to both (iii) inter-subject differences and (iv) experimental effects of interest, as demonstrated by propofol-induced general anaesthesia. Our findings reveal vast and systematic variability across pipelines’ suitability for functional connectomics. Choice of the wrong data-processing pipeline can lead to results that are not only misleading, but systematically so, distorting the functional connectome more drastically than the passage of several months. We also found that the majority of pipelines failed to meet at least one of our criteria. However, we identified 8 candidates satisfying all criteria across each of four independent datasets spanning minutes, weeks, and months, ensuring the generalisability of our recommendations. Our results also generalise to alternative acquisition parameters and preprocessing and denoising choices. By providing the community with a full breakdown of each pipeline’s performance across this multi-dataset, multi-criteria, multi-scale and multi-step approach, we establish a comprehensive set of benchmarks to inform future best practices in functional connectomics.