Feature Matching

The analysis and synthesis components of feature matching are categorized as either landmark or correlation based. Landmark-based feature matching is illustrated for manual global SN in this chapter. Anatomical feature matching is the goal, and corresponding features must be present in source and target brains. In cases where matching features cannot be visually matched, indirect matching methods, such as multi-landmark fitting of the AC-PC line, have been used [23]. Anatomical feature landmarks should have a high degree of penetrance, represent critical areas to fit, and be easily extracted from source and target images. Numerous landmark-based global SN methods have been reported [2,32,38]. Landmark-based manual SN involves human interaction for analysis (feature extraction and comparison) and synthesis (determination of appropriate transform parameters). Pelizzari introduced a surface-based method to automate landmark-based feature matching for intra- and intermodality registration [42]. This surface-based strategy was refined to achieve automated spatial normalization by matching the convex hull of the source brain to a target convex hull derived from the 1988 Talairach Atlas [8,9,35]. Whereas prominent features of the brain, such as its surface, are readily discernible, small, low-contrast features are hard to identify in low-resolution images, limiting the use of a large number of features. Even in highresolution MR images, manual selection of a large number of features can be problematic, restricting the use of landmark-based methods to global or low degree-of-freedom regional spatial normalization. For additional feature matching strategies see the chapter entitled "Warping Strategies for Intersubject Registration".

0 0

Post a comment