Optimization of graph construction can significantly increase the power of structural brain network studies

Abstract

Structural brain networks derived from diffusion magnetic resonance imaging data have been used extensively to describe the human brain, and graph theory has allowed quantification of their network properties. Schemes used to construct the graphs that represent the structural brain networks differ in the metrics they use as edge weights and the algorithms they use to define the network topologies. In this work, twenty graph construction schemes were considered. The schemes use the number of streamlines, the fractional anisotropy, the mean diffusivity or other attributes of the tracts to define the edge weights, and either an absolute threshold or a data-driven algorithm to define the graph topology. The test-retest data of the Human Connectome Project were used to compare the reproducibility of the graphs and their various attributes (edges, topologies, graph theoretical metrics) derived through those schemes, for diffusion images acquired with three different diffusion weightings. The impact of the scheme on the statistical power of the study and on the number of participants required to detect a difference between populations or an effect of an intervention was also calculated. The reproducibility of the graphs and their attributes depended heavily on the graph construction scheme. Graph reproducibility was higher for schemes that used thresholding to define the graph topology, while data-driven schemes performed better at topology reproducibility (mean similarities of 0.962 and 0.984 respectively, for graphs derived from diffusion images with s/mm 2 ). Additionally, schemes that used thresholding resulted in better reproducibility for local graph theoretical metrics (intra-class correlation coefficients (ICC) of the order of 0.8), compared to data-driven schemes. Thresholded and data-driven schemes resulted in high (0.86 or higher) ICCs only for schemes that use exclusively the number of streamlines to construct the graphs. Crucially, the number of participants required to detect a difference between populations or an effect of an intervention could change by a factor of two or more depending on the scheme used, affecting the power of studies to reveal the effects of interest.

Network diffusion modeling predicts neurodegeneration in traumatic brain injury

“…Other definitions of edge weight, such as fractional anisotropy, mean diffusivity, level of myelination, might also be used in further work. 35 Another important limitation of the present study is the relatively small sample size. Notwithstanding, the study provides proof-of-concept to enable the use of similar modeling techniques in larger groups to confirm and extend our results.…”

Combining network topology and information theory to construct representative brain networks

“…Uniquely among the filtering schemes considered here (except for MST-ECO), OMST does not restrict itself to preserving only the strongest edges, and OMST-derived networks are likely to contain a number of edges that all other methods would discard as too weak. The importance of weak connections acting as “shortcuts” to improve efficiency in the brain has been increasingly recognized (Gallos, Makse, & Sigman, 2012 ; Gallos, Sigman, & Makse, 2012 ), and indeed OMST outperformed several other thresholding schemes in terms of both recognition accuracy and reliability of the resulting graph metrics (Dimitriadis, Salis, et al, 2017 ; Messaritaki et al, 2019 ).…”

“…Diffusion MRI is only an indirect measure of in vivo structural connectivity; although correlated with tract-tracing results (Delettre et al, 2019 ; Donahue et al, 2016 ), dMRI tractography also suffers from false positives and difficulties resolving crossing fibers, and not all white matter connections may be correctly identified (Yeh et al, 2018 ). Tractography results may also depend on acquisition and diffusion-weighting protocols (Messaritaki et al, 2019 ). Furthermore, alternative methods for structural network reconstruction such as probabilistic tractography can themselves require the choice of a filtering scheme before binarization—highlighting the key dependence of SDM on deterministic tractography to produce inherently sparse structural networks.…”

“…Existing work comparing aspects of brain network construction has sought to rank pipelines based on reproducibility or test-retest reliability (Andellini, Cannatà, Gazzellini, Bernardi, & Napolitano, 2015 ; Arslan et al, 2018 ; Braun et al, 2012 ; B. Cao et al, 2019 ; H. Cao et al, 2014 ; Du et al, 2015 ; Hu et al, 2019 ; Messaritaki, Dimitriadis, & Jones, 2019 ; Termenon, Jaillard, Delon-Martin, & Achard, 2016 ; J.-H. Wang et al, 2011 ; J. Wang et al, 2017 ; Welton, Kent, Auer, & Dineen, 2015 ), or in terms of maximizing discriminative ability or informativeness of the resulting networks (Arslan et al, 2018 ; De Vico Fallani et al, 2017 ; Dimitriadis, Salis, Tarnanas, & Linden, 2017 ). However, to our knowledge a systematic evaluation of the representativeness of brain networks, in terms of similarity between networks obtained from different construction pipelines applied to the same brain data, has not been undertaken before.…”

Predicting MEG brain functional connectivity using microstructural information

“…In fact, the binarized SC matrices performed better than the FA-weighted ones in many cases. This finding, along with the fact that the FA has been shown to exhibit poor repeatability as an edge-weighting in other datasets (Messaritaki et al (2019b), Messaritaki et al (2019a)) suggest that the FA is of limited relevance when quantifying structural connectivity and trying to relate it to functional connectivity.…”