Structural network disruption markers explain disability in multiple sclerosis

Objective To evaluate whether structural brain network metrics correlate better with clinical impairment and information processing speed in multiple sclerosis (MS) beyond atrophy measures and white matter lesions. Methods This cross-sectional study included 51 healthy controls and 122 patients comprising 58 relapsing–remitting, 28 primary progressive and 36 secondary progressive. Structural brain networks were reconstructed from diffusion-weighted MRIs and standard metrics reflecting network density, efficiency and clustering coefficient were derived and compared between subjects’ groups. Stepwise linear regression analyses were used to investigate the contribution of network measures that explain clinical disability (Expanded Disability Status Scale (EDSS)) and information processing speed (Symbol Digit Modalities Test (SDMT)) compared with conventional MRI metrics alone and to determine the best statistical model that explains better EDSS and SDMT. Results Compared with controls, network efficiency and clustering coefficient were reduced in MS while these measures were also reduced in secondary progressive relative to relapsing–remitting patients. Structural network metrics increase the variance explained by the statistical models for clinical and information processing dysfunction. The best model for EDSS showed that reduced network density and global efficiency and increased age were associated with increased clinical disability. The best model for SDMT showed that lower deep grey matter volume, reduced efficiency and male gender were associated with worse information processing speed. Conclusions Structural topological changes exist between subjects’ groups. Network density and global efficiency explained disability above non-network measures, highlighting that network metrics can provide clinically relevant information about MS pathology.


Registration between T1-weighted and diffusion-weighted images 33
A non-rigid transformation was performed to register the subject's non-filled T1-weighted 34 image to the corresponding diffusion-weighting image (DWI) using BrainSuite [6]. The target 35 volume was the first b=0 image after DWI pre-processing, resulting in a structural image of 36 resolution 2x2x2 mm 3 . The purpose of registering the structural images to the diffusion images 37 at this stage is two-fold: a) matching the voxel dimensions and positions of the T1-scan to that 38 of DWI means that any subsequent image derived from the anatomical scan will be inherently 39 aligned to the DWI; and b) aligning the anatomical image to the DWI and not the other way 40 around ensures that a re-orientation of the gradient direction is not required. 41

Tissue segmentation and parcellation 42
We non-rigidly transformed the lesions to DWI space and then filled the T1-weighted images 43 in this space using a modality-agnostic patch-based method [7]. The reason that we registered 44 we registered the T1-image in DWI space before lesion filling so that we matched all the 45 anatomical features between the two modalities incusing lesions. Hence, we ensured that the 46 non-rigid registration was not affected by the lesion filling. The filled T1-weighted images 47 were then segmented into cortical grey matter, white matter, deep grey matter, brainstem and 48 cerebrospinal fluid (CSF) and parcellated into anatomically distinct regions according to 49 Desikan-Killiany-Tourville atlas protocol using the GIF framework [8]. This method has been 50 previously used in different neurological diseases such as MS [9], dementia [10] and epilepsy 51 [11] GIF is freely available as web-service at http://cmictig.cs.ucl.ac.uk/niftyweb [12]. We then 52 estimated the volumes of the various tissue types (NABV (normal appearing brain volume 53 (BV)), GM, CGM (cortical GM), DGM (deep GM)). Reduction of these volumes reflects 54 atrophy. LL (lesion load) was also computed as a measure of WM focal damage. 55 56 D. Diffusion-weighted imaging processing

B0 registration, eddy current and susceptibility induced correction 90
The mean b0 image was rigid registered to the first b0 image. Then, the same rigid 91 transformation was applied to the 61 DWI volumes. FSL v5.0.9 was used on the DWI data to 92 correct for eddy current and head motion [13]. We also corrected for susceptibility induced 93 distortions caused by EPI sequences using BrainSuite v.15b. This method uses the T1-weighted 94 image as the registration-template to correct the diffusion data [6]. 95

Model response function and Constrained Spherical Deconvolution 96
For the subsequent steps, we used MRtrix3 v0.3.14. We estimated the response function [14], 97 the signal expected from a voxel that contains a single coherent fibre bundle, and then we 98 performed constrained spherical deconvolution (CSD) [15,16] to estimate the voxel-wise fibre 99 orientation distribution (FOD). 100

Whole-brain streamline tractography 101
For each subject, 10 7 streamlines were generated. For the probabilistic tractography, the iFOD2 102 algorithm [17] was employed using the default parameters -step size=1.25 mm, maximum length=250 mm, implementing the anatomically constrained tractography (ACT) framework 104 [18]. Spherical-deconvolution informed filtering of tractograms (SIFT2) was applied to the 105 generated tractograms to modulate the contribution of each streamline to the relevant edge [19]. 106 In this way the streamline count is reflective of the underlying fibre density at the local level. 107 When looking at the connection density of a particular pathway, this interpretation remains 108 such that a larger region is likely to be intersected by a greater number of streamlines. In fact, 109 Yeh,Smith [20] showed that the application of ACT and SIFT2 (both techniques were also 110 applied in our study) improves the biological accuracy of the reconstructed connectome while 111 other scaling methods provide only incomplete correction. 112 113 E. Network reconstruction 114 GM parcellations constituted the network nodes, 120 in total. Each network edge was defined 115 as the sum of weights of streamlines connecting a pair of nodes [19]. The pipeline is 116 summarised in Fig. 1. To assess the network topology, we extracted the following network 117 measures: 118 Edge Density: also known as connectivity, this is defined as the percentage of connections that 119 exist relative to the potential number of network connections [21]. 120 Global efficiency: is the average of the inverse of the distance matrix of the entire network 121 matrix [22]. It is a measure of the overall information transfer efficiency across the whole 122 network. 123 Local efficiency: similar to global efficiency, it is defined as the average of the inverse distance 124 matrix but in a sub-cluster of the network [22]. It is considered as a measure of the local 125 information flow. As this is a node-specific measure we average over all the nodes to get the 126 mean local efficiency metric. 127 Clustering coefficient: is also a node-specific measure which describes local organisation 128 reflecting the number of connections between the neighbours of each node [23]. Averaging 129 over all the nodes provides the mean clustering coefficient. 130 The metrics were derived using the TractoR [24]