The Diffusion-Simulated Connectivity Dataset

Published: 08-04-2021| Version 1 | DOI: 10.17632/fgf86jdfg6.1
Contributors:
Jonathan Rafael-Patino,
Raphael Truffet,
Jean-Philippe Thiran

Description

The Diffusion-Simulated Connectivity Dataset is generated from a numerical phantom and Monte-Carlo simulation of the diffusion-weighted MRI signal [Rafael-Patino et al., Frontiers in Neuroinformatics, 2020]. The dataset was created to evaluate diffusion-weighted MRI-based structural connectome estimation techniques such as voxel-wise fibre orientation estimation, tractography, and structural connectome estimation. The phantom is constructed from 12,196 synthetic tubular fibres (strands) ranging in diameter from 1.4μm to 4.2μm. The phantom geometry was obtained using the optimisation procedure included in the Numerical Fiber Generator [Close et al., NeuroImage, 2009]. Sixteen regions of interest (ROIs) located on the surface of a sphere are connected through a variable number of strands. The strands form complex white matter configurations that include kissing, branching, and crossing at different angles. The weights of the 16x16 ground truth connectivity matrix are defined by the total cross-sectional area of the strands forming the connections. The phantom has a volume of 1 cubic millimetre and is divided into a 40x40x40 voxel image matrix. After the Monte-Carlo simulation of the diffusion-weighted MRI signals, the image header is set to a voxel size of 1.0 mm isotropic, creating a final image size of 4x4x4 cm³, compatible with conventional diffusion-weighted MRI methods. The diffusion-weighted MRI simulation protocol includes 360 diffusion-weighted images and 4 non-diffusion-weighted images (b=0 s/mm²). The diffusion-weighted MRI signals are distributed on 4 b-shells (b=1000, 1925, 3094, 13191 s/mm²). The dataset includes the signal generated with the Monte Carlo simulation and the same signal corrupted with various levels of Rician noise. Additionally, the dataset contains the relevant phantom’s ground-truth information including, i) a voxel map indicating the volume fraction occupied by the strands, ii) a label map of the location of the 16 ROIs, iii) the 12,196 strands' centerline trajectories, as well as their diameter and ending ROIs, and iv) the three-dimensional substrate mesh used for the Monte Carlo simulation.

Files