attributed to a single tissue class without ambiguity. Sometimes a codebook vector is positioned at the border between two tissue classes, which makes a clear assignment difficult or even impossible. In these situations, a hard, i.e., unique assignment of the codebook vector to a single tissue class leads to misclassifications resulting in decreased segmentation quality.
Detailed segmentation results using vector quantization are listed in Section 9.3.
In the following, the results for classification by a GRBF network after preceding vector quantization of the gray-level feature space are discussed. The parameters of the GRBF classifier are listed Table 3. The training data comprise approximately 1% of the whole data set each. Detailed numbers are listed in Tables 4 and 5.
This approach yields better segmentation results than unsupervised clustering with respect to subjective evaluation by a neuroradiologist, as can be seen from Table 2. Typical segmentation results are presented in Figs 9c, 10c, and 11c.
The improvement of segmentation quality, however, was accompanied by a considerable increase of human intervention: Manual processing time for interactive labeling of the training data was approximately 30 minutes per data set. Thus, supervised classification requires more human intervention than unsupervised clustering.
Table 6 presents a detailed list of segmentation results using GRBF classification.
Tables 7 through 11 present a comparative evaluation of results obtained by the two different segmentation approaches using contingency tables. On average, 92.6% of the voxels were assigned to corresponding class labels.
Was this article helpful?