Non-Negative Data-Driven Mapping of Structural Connections in the Neonatal Brain

Abstract

Mapping connections in the neonatal brain can provide insight into the crucial early stages of neurodevelopment that shape brain organisation and lay the foundations for cognition and behaviour. Diffusion MRI and tractography provide unique opportunities for such explorations, through estimation of white matter bundles and brain connectivity. Atlas-based tractography protocols, i.e. a priori defined sets of masks and logical operations in a template space, have been commonly used in the adult brain to drive such explorations. However, rapid growth and maturation of the brain during early development make it challenging to ensure correspondence and validity of such atlas-based tractography approaches in the developing brain. An alternative can be provided by data-driven methods, which do not depend on predefined regions of interest. Here, we develop a novel data-driven framework to extract white matter bundles and their associated grey matter networks from neonatal tractography data, based on non-negative matrix factorisation that is inherently suited to the non-negative nature of structural connectivity data. We also develop a non-negative dual regression framework to map group-level components to individual subjects. Using in-silico simulations, we evaluate the accuracy of our approach in extracting connectivity components and compare with an alternative data-driven method, independent component analysis. We apply non-negative matrix factorisation to whole-brain connectivity obtained from publicly available datasets from the Developing Human Connectome Project, yielding grey matter components and their corresponding white matter bundles. We assess the validity and interpretability of these components against traditional tractography results and grey matter networks obtained from resting-state fMRI in the same subjects. We subsequently use them to generate a parcellation of the neonatal cortex using data from 323 new-born babies and we assess the robustness and reproducibility of this connectivity-driven parcellation.

Non-negative data-driven mapping of structural connections with application to the neonatal brain

“…It has also been shown that tractography streamlines are biased towards terminations in the gyri rather than the sulci ( Schilling et al, 2018 ; Van Essen et al, 2013 ), although the effects of this “gyral bias” can be minimised by seeding from the cortical surface rather than the whole brain ( Donahue et al, 2016 ; Schilling et al, 2018 ), as we have done here. We have also shown in previous work that the effects of gyral bias are less prevalent in neonates than in adults due to the less developed cortical folding ( Thompson et al, 2019 )and we therefore expect less direct influence of such biases into the NMF performance in the neonatal brain. In fact, our parcellation borders did not show a consistent overlap with sulcal fundi or gyral crowns ( Suppl.…”

“…We subsequently seeded 10,000 streamlines from each of 58,551 vertices on the WGB of both hemispheres (average vertex spacing 1.2 mm, excluding the medial wall) and from each of 2548 subcortical 2mm voxels (bilateral amygdala, caudate, hippocampus, putamen and thalamus), giving us a total of N = 61,099 seeds. This type of grey matter seeding has been shown to suffer less from the gyral bias in tractography, compared to whole-brainwhite matter seeding, even if gyral bias is less prominent in the neonatal brain ( Thompson et al, 2019 ). Visitation counts were recorded between each seed point and each of M = 50,272 voxels in a whole-brain mask with the ventricles removed, down-sampled to 2 mm 3 .…”

Multiplex dynamic networks in the newborn brain disclose latent links with neurobehavioral phenotypes

“…We extracted a latent representation from the group-level tensor using non-negative matrix factorization (NMF); a low-rank matrix approximation tool that is appropriate for non-negative data such as cortico-cortical interactions (Mahyari et al, 2017; (Stage 3 in Figure 1). The NMF technique was selected because it provides robust reconstruction results when compared to PCA and ICA (Thompson et al, 2020;, and because of its effectiveness in suppressing intersubject variability (Calesella et al, 2020;Calesella et al, 2021). We applied the NMF to the group tensor's frequency slices and the resultant estimates were aggregated back to their original tensor form (see Supplementary Section 2).…”