7.1 VA Reading - Smith 2006 NeuroImage (cont.)

Smith SM, Jenkinson M, Johansen-Berg H, Rueckert D, Nichols TE, Mackay CE, Watkins KE, Ciccarelli O, Cader MZ, Matthews PM, Behrens TEJ (2006). Tract-based spatial statistics: Voxelwise analysis of multi-subject diffusion data. NeuroImage; 31: 1487-505. 


Creating the mean FA image and its skeleton

• all subjects' FA images are aligned to the most typical subject
• affine-transform the entire aligned dataset into 1*1*1 mm^3 MNI152 space (easy to interpret and display)
• choose higher resolution to avoid partial volume effect and interpolation blurring; too high will result in slower computation and unnecessarily large data files
• a mean FA image is created; relatively smooth locally because of averaging as well as resolution upsampling.
• the mean FA is fed into the tract skeleton generation, which represent all tracts with are "common" to all subjects.
• corpus callosum is sheet-like, therefore the skeleton should run along the center of the sheet
• cingulum bundle is tube-like, , therefore the skeleton should run along the center of the tube
• estimate the local surface perpendicular direction and then perform non-maximum-suppression, i.e. a voxel with highest FA in perpendicular direction of the tract is identified as the centre of the tract.
• Local FA centre-of gravity is used in the first stage of skeletonization
• FA image second derivative is used in the second stage of skeletonization
  
• If the voxel of interest lies away from a tract centre, FA will be higher in the neighbouring voxels on one side of the voxel than on the other - the direction in which it is highest points towards the nearest tract centre (this took me a while to digest).
• regularise the estimated tract perpendicular direction in order to improve estimation robustness
• search for the centre of each tract; if the FA value is greater than the neighbouring values, then the voxel is marked as lying on the skeleton.
• The FA skeleton need to be thresholded to restrict further analysis to points which are within white matter which has been successfully aligned across subjects.
• thresholding between 0.2 and 0.3 successfully excludes voxels which are primarily grey matter or CSF in the majority of subjects
• Skeleton may tend to be disconnected because the tract perpendicular direction is not well-defined at junctions
• a more sophisticated projection method specifically for junctions needs to be developed