Automatic Segmentation of the Caudate Nuclei using Active Appearance Models
Kola Babalola, V. S. Petrovic, Adam Mills · 2007
Abstract. We describe the application of an active appearance model (AAM) based method to segmentation of the caudate nuclei. A “composite” 3D profile AAM was constructed from the surfaces of 15 subcortical structures using a training set of 50 subjects, and individual AAMs of the left and right caudate constructed from 227 subjects. Segmentation starts with affine registration to initialise the composite model within the image, then a search using the composite model. This provides a reliable but coarse segmentation, used to initialise search with the individual caudate models. The results are further refined by reclassifying voxels within a one-voxel neighbourhood of the surface. Application to 24 subjects resulted in 2 failed searches (L+R of same subject) and a mean Tanimoto overlap of 73.1 % (failed searches excluded). The overall official score (failed searches included) was 71. The coefficient of variation when applied to 10 independently acquired datasets of the same subject was 3.5 % (1.9 % for left and 5.1 % for right). 1