Visit DIX: German-Spanish-German dictionary | diccionario Alemán-Castellano-Alemán | Spanisch-Deutsch-Spanisch Wörterbuch |
The classification of tensor data should connect regions with similar local structure, i.e. fibertracts in the corpus callosum. These fibertracts should be present in all subjects. An alignment of tensor data should match the corpus callosum of two subjects. As explained in the previous section the classification would also allow to test the alignment process by comparing the classified data sets.
One possible way of classification of tensor data has already been presented in Section 4.3 where the tensors have been classified depending on their shape.
This is the simplest approach where the classification is based only on the values of the voxel itself. The classification and the measurements are best based on the eigenfield, since the eigen-domain has a much more intuitive meaning. If a useful classification has been found, it is worthwhile to find a corresponding one in the tensor-domain, since this would save the eigensystem decomposition of the tensorfield.
Different measurements can lead to a number of different classifications. Examples are:
Any tensor that is not assigned a label but is not equal to zero is assigned a ''rest-group'' label so that the body does not get mixed with the image background. Finally a histogram of the labeled image is build and the labels are sorted according to their size. The resulting labels can be displayed with the program main.
The described process is certainly in a very early stage and does not produce useful results yet. But it helps to get an insight on the tensor data and can be used as a starting point for further implementations.
The label-growing approach has also been used to ''clean'' the mask generated in Section 6.1. The thresholded T2 weighted image is labeled, (with the threshold for the neighborhood set to zero), so that any connected group is assigned the same label. Then only the largest label is kept, so that the brain, as the largest object in the image, is the only remaining label.
Segmentation into different tissue classes can be performed using imager8.1. This program allows to manually select example points of several images and does a segmentation of the data based on k-Nearest-Neighbor estimation.
The following Figure 8.1 shows some results obtained with the implemented classification method.
|