Is removal of weak connections necessary for graph-theoretical analysis of dense weighted structural connectomes from diffusion MRI?

Abstract

Recent advances in diffusion MRI tractography permit the generation of dense weighted structural connectomes that offer greater insight into brain organization. However, these efforts are hampered by the lack of consensus on how to extract topological measures from the resulting graphs. Here we evaluate the common practice of removing the graphs' weak connections, which is primarily intended to eliminate spurious connections and emphasize strong connections. Because this processing step requires arbitrary or heuristic-based choices (e.g., setting a threshold level below which connections are removed), and such choices might complicate statistical analysis and inter-study comparisons, in this work we test whether removing weak connections is indeed necessary. To this end, we systematically evaluated the effect of removing weak connections on a range of popular graphtheoretical metrics. Specifically, we investigated if (and at what extent) removal of weak connections introduces a statistically significant difference between two otherwise equal groups of healthy subjects when only applied to one of the groups. Using data from the Human Connectome Project, we found that removal of weak connections had no statistical effect even when removing the weakest ~70-90% connections. Removing yet a larger extent of weak connections, thus reducing connectivity density even further, did produce a predictably significant effect. However, metric values became sensitive to the exact connectivity density, which has ramifications regarding the stability of the statistical analysis. This pattern persisted whether connections were removed by connection strength threshold or connectivity density, and for connectomes generated using parcellations at different resolutions. Finally, we showed that the same pattern also applies for data from a clinical-grade MRI scanner. In conclusion, our analysis revealed that removing weak connections is not necessary for graph-theoretical analysis of dense weighted connectomes. Because removal of weak connections provides no practical utility to offset the undesirable requirement for arbitrary or heuristic-based choices, we recommend that this step is avoided in future studies.

Alpha Band Resting-State EEG Connectivity Is Associated With Non-verbal Intelligence

“…Our results are in line with recent diffusion tensor imaging study, where the preserving of weak connections on calculation of graph connectivity metrics was advocated for Civier et al (2019), and with earlier fMRI study where thresholding has been shown to lead to inconsistent results (Garrison et al, 2015). The results suggest that common practice in research to eliminate weak connections may lead to missing important information.…”

The effect of network thresholding and weighting on structural brain networks in the UK Biobank

“…In combination with results from multiple approaches indicating pervasive degeneration of the brain's white matter connections with adult age (Burzynska et al, 2010;Cox et al, 2016;Salat, 2011), our finding that the strongest age-associations fell between 50% and 95% network sparsity, and that the profile of age-associations within the discarded connections were predominantly null across sparsities is consistent with: 1) prior evidence that the human brain is likely to have a connection density of ~30% (Roberts et al, 2017); 2) that probabilistic tractography fundamentally over-estimates the number of connections (Roberts et al, 2017), and; 3) that these spurious connections add noise to the signal (Jbabdi and Johansen-Berg, 2011). Moreover, our results are inconsistent with previous reports that topological network properties of connectomes are not significantly altered by the removal of weak connections (Civier et al, 2019).…”

“…A recent population-based atlas of white matter connectivity, constructed by manually labelling 40 major streamline clusters (Yeh et al, 2018), shows promise in validating connectivity but is limited to a small proportion of possible network connections.In the absence of a comprehensive map of connectivity and to meet a demand for more principled network denoising approaches (de Reus and van den Heuvel, 2013;Maier-Hein et al, 2017;Van Wijk et al, 2010), researchers have introduced inferential methods to identify and discard potentially spurious connections. Some researchers have advocated using raw (unthresholded) matrices without removal of any connections on the basis that topological network properties are not significantly altered by the inclusion of weak connections (Civier et al, 2019). However, many network studies have employed thresholding strategies, such as, absolute-thresholding which applies a uniform threshold to retain only connections above a set weight (Hagmann et al, 2007), and density-thresholding which applies a (relative) threshold on the connection weights such that the weakest connections are removed to match the same number of connections across subjects (Rubinov and Sporns, 2010).…”

Network communication models improve the behavioral and functional predictive utility of the human structural connectome

“…Although we explored multiple brain network reconstruction pipelines, we were not primarily concerned with which mapping techniques produced connectomes with the highest predictive utility. The choice of parcellation schemes [70] and whether or not to threshold structural connectomes [71,72] are both complex open questions that fall outside the scope of this work. We also note that white matter tractography algorithms are susceptible to a number of known biases that could potentially impact our results [73].…”