Home Computer Science Computational Diffusion MRI: MICCAI Workshop, Athens, Greece, October 2016
In order to utilise the high resolution (0.7 mm isotropic) of the available structural data we chose the HCP FreeSurfer pipeline over the standard recon-all pipeline to generate cortical surface reconstructions for each subject. This improved pipeline does not down-sample the T1w images to 1 mm isotropic resolution, and incorporates the additional information available in the T2w scans to reduce surface placement errors .
Following cortical surface reconstruction the diffusion datasets of each subject were sampled at the midpoint between the white/grey matter (WM/GM) boundary and the pial surface so as to reduce the likelihood of either WM or CSF contamination [20, 22].
In a similar procedure to that of Nagy et al. , a sixth order spherical harmonic series (SHS) was fit to the dMRI signal in order to characterise the apparent diffusion coefficient (ADC) profile of the cortical tissue. A SHS was fit separately in each b- shell for each surface vertex of the right hemisphere of each subject. A subset of the features presented in  were calculated from the ADC to obtain a [1x5] feature vector per vertex, per b-shell. The features, as detailed below, characterise the ADC profile in relation to the local surface normal, and therefore describe the GM tissue irrespective of the orientation differences that result from cortical folding.
1. The value of the ADC profile along the surface normal.
2. The mean of the ADC profile in the plane perpendicular to the surface normal, i.e. parallel to the cortical sheet. C(n) is the unit circle perpendicular to n.
3-5. The k=2,3 and 4 moments, respectively, of the ADC in the plane perpendicular to the surface normal.
The group average of each of the 15 features was computed in turn using sulcus- based surface averaging . This approach allows an individual’s folding pattern to be aligned to an average folding pattern, in this case on the fsaverage surface. With this method, any given fsaverage vertex will combine data from individual subject vertices that have surface normals in different directions. This makes it possible to detect local-surface-geometry-dependent diffusion signatures of cortical areas even though their local normal directions might differ from subject to subject. This information would be compromised if the diffusion data were to be directly averaged in 3D (folded) space. The transformation between each subject’s cortical surface and the target brain space was applied to each of the cortical features in turn. The mean across all subjects of each feature was then calculated for each vertex of the fsaverage surface. Finally, the averaged features were recombined into a [1x15] group average feature space for classification.
|< Prev||CONTENTS||Next >|