6. Enhancing Contours Features

Blurred features introduce ambiguities during landmark identiication and placement which can account for an increase in local distortions during the registration procedure. To minimise placement errors three enhancement procedures based on spatial competition, feature selection and feature thresholding are proposed and experimentally investigated. 6.1 Feature detection and representation In chapter 5 a multi-layered, self-organising architecture was employed to extract landmarks and position them at organ boundaries. Using vector quantisation, it is possible to develop characteristic contour features representing statistically signiicant properties of the data. However , many features lack contour deenition, ie. considerably large amounts of blurring exist. While it is possible to reduce the amount of blurring by increasing the feature set size, the redundancy of the set increases correspondingly, causing feature matching diiculties as well as a reduction in processing speed. Blurring can contribute to ambiguous landmark identiication and positioning. This is illustrated in gure 6.1. The contour of a test pattern is given in (a). The top left corner and the bottom right corner act as target landmark positions. The contour of the latter has been corrupted to illustrate the eeect of noise. Figures (b) and (c) show the internal representations of two corner feature detectors. While (b) typiies a sharp, well deened instance, (c) acts as its blurred counterpart. The output of the detector plane of (b) is given in (d). Two well deened peaks are visible at the top left corner and bottom right corner. Even though less distinct, the distorted corner pattern can still be identiied with conndence. The output of the detector plane of (c) is shown in (e). The entire surface is smoother than (d). The uncorrupted corner and the blurred feature give a less pronounced peak; the position of the corrupted corner cannot be detected with conndence and several likely locations are indicated by the smooth hill. Thus, detection and placement can be improved by using sharp feature representations. The aim of this chapter is to develop feature sets with sharp contours. Three amendments to the previously proposed architecture are proposed: the use of spatial competition during training is outlined in x6.2, the selection of a subset of features from a larger set is suggested in x6.3, and the application of threshold-like, feature post-processing is discussed in x6.4. First a description of the three methods is given which is followed by an experimental investigation in x6.5. The new feature types of …