Info

"The quality of each image of a 3D data set was evaluated by a neuroradiologist on a 5-grade semiquantitative score (1 = excellent, 5 = insufficient). The table contains the average scores of all the images belonging to a 3D data set.

"The quality of each image of a 3D data set was evaluated by a neuroradiologist on a 5-grade semiquantitative score (1 = excellent, 5 = insufficient). The table contains the average scores of all the images belonging to a 3D data set.

cluster-specific centroids (x) of the data set. This tesselation of the feature space provides the basis for two methods of tissue segmentation that will be explained in Section 8.

It should be mentioned that for reducing the computational expense of the procedure, it is useful not to present the total data set X, but to restrict the input of the vector quantizer to a representative subset X' c X of randomly chosen feature vectors x' e X'.

8 Classification

Given a set of feature vectors x in a gray-level feature space G, vector quantization can determine a set of prototypical codebook vectors Wj representing the feature space. This provides the basis for segmentation of the imaging data set with respect to different tissue classes. Two alternative approaches are discussed in the following paragraphs.

8.1 Interactive Assignment of Codebook Vectors

This approach requires two steps: In a first step, each feature vector x is uniquely attributed to a codebook vector Wj according to a minimal distance criterion. If a vector x is presented as an input to a GRBF network sketched in Fig. 1, the N neurons of the hidden layer are activated according to

TABLE 3 Parameters employed for vector quantization of the gray level feature space (see Section 7) and GRBF classification (see Section 8.2)

# Codebook vectors N Vq # Iteration steps tmax

# Classes m

0 0

Post a comment