Classification

The parameters for dark rim and focal abnormality were measured from a set of images identified by radiologists and confirmed by histology. Two sets each were used for thick, thin, and ruptured caps. Several measurements along the contour were thus available for each set. The mean and covariance of each parameter for thick, thin, and ruptured caps was then calculated. These templates were used for classification by the feature distance of a candidate point from a template for thick, thin, and ruptured classes. The Mahalanobis distance of the dark rim parameter was used to differentiate thick caps from the other two classes. The thin and ruptured classes were differentiated from the remaining points based on the curvature parameter again using the Mahalanobis distance metric, r2 = (x - m)C-1(x - m) (8.61)

where m and C are mean and covariance matrices, respectively. This decision is based on the observation that both thin and ruptured caps do not have a dark rim but the ruptured can be differentiated by the presence of a focal contour abnormality. Figure 8.27 shows an example of the algorithms classification compared to ground truth by histology.

' ■ ' * ^fâjïfr r T1

Segments by histology

Segments by algorithm

Figure 8.27: Example of FC classification with corresponding MR images and ground truth by histology. (A color version of this figure can be found on the CD. Green: thick cap; blue: thin cap; red: ruptured cap.)

0 0

Post a comment