07-11-2012, 12:36 PM
Automatic Segmentation of Human Brain Tractography Data
Automatic Segmentation.pdf (Size: 3.23 MB / Downloads: 29)
DTI
Diusion of water molecule probes tissue structure at the smallest i.e.
voxel resolution.
Water diusion in oriented brous structures (as white matter and
muscles) is anisotropic (varies with direction).
Tensors are used to model diusion. Major eigen vectors of the tensor
(Principle Directions) gives the ber tract direction.
Tractography
Streamline Tractography estimates white matter tract trajectories
following the most likely tract direction. It locally chooses the most
likely ber trajectory.
Entire brain tractography is estimated by stepping along the major
eigen vector direction using RK method.
Introduction
A ber is represented as a set of points in 3D space (typically 20 to
30 points per ber).
Tractography produces thousands of ber trajectories per subject
(250K).
Tractography can be seen as a 3D point cloud but that is not very
useful.
Useful information can be extracted only when they are organized into
anatomically meaningful structure.
Atlas creation
Learn common white matter structure present across the subjects and
build white matter atlas.
Project the bers into a high dimensional space where each ber can
be seen as a point.
Cluster the points in that space using any central clustering technique
(as K-mean or GMM).
Label the clusters formed using experts knowledge (allot anatomical
names).
This constitutes the anatomical model of white matter structure
(High Dimensional White Matter Atlas).
Automatic Atlas Based
Madah [16] proposed a manual interaction method to create a
tractography atlas and gave an algorithm for transferring its labels to
novel subjects.
Madah [17] further proposed an atlas creation method for corpus
callosum using labeled tractography from several subjects and then
used Expectation Maximization (EM) framework to classify bers of
novel subjects.
Automatic Segmentation
Atlas information can be used to segment novel brain. New tractography
path is embedded in the same space in which the clustering was performed
originally. Cluster labels and anatomical informations are assigned
according to the nearest cluster centroid.
Classication and Labeling
New subject's embedding vectors are labelled according to the nearest
cluster centroid, giving a cluster label for each path.
Per cluster anatomical label and any addition (say color) informations
are transferred to the novel subject