Proceedings of the Symposium on Big Data Initiatives for Connectomic Research 2015 International conference on Brain Informatics and Health

9 A pipeline for automatic semantic annotation of human connectomics revealed by diffusion tractography Introduction Predictive modelling of functional connectomes from structural connectomes explores commonalities across multi-modal imaging data [1]. Brain structure not only constrains but also shapes functional connectomes. We exploit this property to devise a statistically sound way, based on a connectivity identification framework. This applies the Randomised Lasso (RL) principle to sparse Canonical Correlation Analysis (sCCA) [2], and transports functional connectomes on a common Riemannian manifold to retain their symmetric positive definite (SPD) geometry [3]. This framework highlights the structural connections that are consistently selected in predicting functional connectomes, across microstructural indices. Fractional anisotropy (FA) and mean diffusivity (MD) are derived from the conventional tensor model, whereas intra-cellular volume fraction (ICVF), orientation dispersion index (ODI), the concentration parameter (Kappa) and the volume fraction of the isotropic compartment (ISO) are derived from a more biophysically plausible model based on Neurite Orientation Dispersion and Density Imaging (NODDI) [4]. Methods NODDI data were obtained from 19 healthy volunteers on a 1.5T Siemens scanner based on three shells of Diffusion Weighted (DW)-MRI with b=2400smm-2, b=800smm-2 and b=300smm-2 were acquired with 240x240x150mm FOV, voxel size of 2.5x2.5x2.5mm and TR/TE=8300/98ms. Resting-state (rs)-fMRI were acquired: TR/TE=2160/30msec, effective voxel size 4.03.33.3mm, FOV 210210120mm. We run probabilistic tractography on the data acquired with b=2400smm-2, and we obtain the streamlines that connect each pair of cortical FreeSurfer regions. For each pair of regions we average the microstructural indices to obtain brain connectomes based on number of streamlines (NSTREAMS) and weighted averages of FA, MD, ICVF, ODI, Kappa and ISO, fig.1. Functional connectivity matrices are estimated as the normalised inverse of the covariance of the averaged fMRI time-series within each region, fig.2. We learn the relationship between microstructural indices (X) and rs-fMRI (Y) across subjects based on sCCA. sCCA extracts sparse vectors that are multiplied by X and Y 1 to maximize their linear relationship. sCCA is applied on vectorised versions of functional and structural connectivity across subjects. However, there is no guarantee that the results from linear operations on the elements of a SPD matrix will also lie on an SPD manifold. A solution is to project each functional con-nectome into a common tangent space (SPD-sCCA), which is constructed based on the average precision matrix. To alleviate the dependence of the extracted connections on the regularisation parameter we use RL [5]. Results Based on leave-one-out cross-validation, we …