We introduce a segmentation approach for brain MR images focusing on the following subcortical brain structures: caudate nucleus, thalamus, putamen, globus pallidus, thalamus, lateral ventricle and third ventricle. The overall image registration and segmentation approach consists of two phases. The first phase is registration of MR images to a standard coordinate and segmentation of brain into three types of tissue. In the second phase, an improved model – based segmentation method for subcortical structures of interest is applied to tissue segmentation result using a set of manually segmented structures as models. The approach first uses the model voxel centroid to locate the structure location, and the distances between valid boundary voxels of both model and incoming image are calculated to perform the structure model selection and matching operations, and finally the model boundary is used to bridge or fill the boundary to obtain the final segmentation result. An improved algorithm introducing direction property to distance calculation in model selection is proposed. A different segmentation approach is applied to lateral ventricle.
Experiments to compare the performance of the original and modified algorithm are carried out. An extensive validation process has been performed with positive results on 11 structures as well. DICE similarity index and modified DICE criterion are used as performance evaluators. The approach is then applied to volumetric study of 11 structures in 113 healthy subjects. The age effect on these structures is investigated and volumes of caudate nucleus, thalamus as well as putamen decrease with aging.