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

NeuroImage

Ian J. Deary¹  David C. Liewald²  Mark E. Bastin³  Stuart J. Ritchie⁴  Elliot M. Tucker-Drob⁴  Colin R. Buchanan⁵  James W. Madole⁵  Simon R. Cox⁶ 

[1] Brain Research Imaging Centre, Neuroimaging Sciences, The University of Edinburgh, Edinburgh, UK;Corresponding author. Lothian Birth Cohorts group, The University of Edinburgh, Edinburgh, EH8 9JZ, UK.;Department of Psychology, The University of Edinburgh, Edinburgh, UK;Scottish Imaging Network, A Platform for Scientific Excellence (SINAPSE) Collaboration, Edinburgh, UK;Lothian Birth Cohorts group, The University of Edinburgh, Edinburgh, UK;Social, Genetic and Developmental Psychiatry Centre, King’s College London, London, UK;

关键词: Aging;    Brain network;    Connectome;    Diffusion MRI;    Thresholding;   

DOI  :  

来源: DOAJ

【 摘 要 】

Whole-brain structural networks can be constructed using diffusion MRI and probabilistic tractography. However, measurement noise and the probabilistic nature of the tracking procedure result in an unknown proportion of spurious white matter connections. Faithful disentanglement of spurious and genuine connections is hindered by a lack of comprehensive anatomical information at the network-level. Therefore, network thresholding methods are widely used to remove ostensibly false connections, but it is not yet clear how different thresholding strategies affect basic network properties and their associations with meaningful demographic variables, such as age. In a sample of 3153 generally healthy volunteers from the UK Biobank Imaging Study (aged 44–77 years), we constructed whole-brain structural networks and applied two principled network thresholding approaches (consistency and proportional thresholding). These were applied over a broad range of threshold levels across six alternative network weightings (streamline count, fractional anisotropy, mean diffusivity and three novel weightings from neurite orientation dispersion and density imaging) and for four common network measures (mean edge weight, characteristic path length, network efficiency and network clustering coefficient). We compared network measures against age associations and found that: 1) measures derived from unthresholded matrices yielded the weakest age-associations (0.033 ​≤ ​|β| ​≤ ​0.409); and 2) the most commonly-used level of proportional-thresholding from the literature (retaining 68.7% of all possible connections) yielded significantly weaker age-associations (0.070 ​≤ ​|β| ​≤ ​0.406) than the consistency-based approach which retained only 30% of connections (0.140 ​≤ ​|β| ​≤ ​0.409). However, we determined that the stringency of the threshold was a stronger determinant of the network-age association than the choice of threshold method and the two thresholding approaches identified a highly overlapping set of connections (ICC ​= ​0.84), when matched at 70% network sparsity. Generally, more stringent thresholding resulted in more age-sensitive network measures in five of the six network weightings, except at the highest levels of sparsity (>90%), where crucial connections were then removed. At two commonly-used threshold levels, the age-associations of the connections that were discarded (mean β ​≤ ​|0.068|) were significantly smaller in magnitude than the corresponding age-associations of the connections that were retained (mean β ​≤ ​|0.219|, p ​< ​0.001, uncorrected). Given histological evidence of widespread degeneration of structural brain connectivity with increasing age, these results indicate that stringent thresholding methods may be most accurate in identifying true white matter connections.

【 授权许可】

Unknown