Model based segmentation methods and their application to biomedical image analysis
Xiaolei Huang, Tian Shen · 2011
Deformable models have been a classic approach to image segmentation, and they have been used with considerable success, e.g. in segmenting medical images. While traditional deformable models rely on local image gradient information, it is important to incorporate additional contextual constraints, in order to achieve robustness to noise, intensity inhomogeneity, and locally weak boundary. In this thesis, we propose three new deformable model based segmentation methods to extract contour, surface and tubular structures from 2D and 3D medical images. First, 2D and 3D Active Volume Models (AVM) are proposed, which can integrate region information. To further improve the robustness and accuracy of AVM, we propose Multiple-Surface Active Volume Model (MSAVM), which is able to incorporate high-level geometric spatial constraints among multiple objects. Second, to segment 3D objects containing complex surfaces or high curvature regions, we propose a 3D Laplacian-driven parametric deformable model, which preserves model mesh quality by using a novel internal force derived from mesh Laplacian. Third, to segment coronary arteries, we propose 4D parametric deformable curves aiming to extract the vessels' centerlines and corresponding radii simultaneously. We present both qualitative and quantitative validation experiments for all three proposed methods.