Location: MIC-Arena, MFC II, Maria-Goeppert-Str. 1a, 23562 Lübeck
Title: Omnidirectional Displacements for Deformable Surfaces
Abstract: Deformable surface meshes are widely used in model-based segmentation of medical images. The basic idea is to deform a given (template) shape in such a way that the deformed shape provides an optimal geometric representation of the corresponding target structure in the image. Usually, the deformable mesh "probes" the image information at each vertex position. Given these probes, a new shape is computed by displacing the vertices of the mesh, following a trade-off between image fidelity and anatomically plausible deformation. Most commonly, one-dimensional image probes (e.g. in surface normal direction) are taken at each vertex of the deformable mesh. This allows for fast and easily regularized mesh deformations. However, in case of high mesh curvature, two fundamental problems arise:
(1) One-dimensional probes are prone to miss features that the target structure exhibits in the image data.
(2) One-dimensional probes frequently detect "wrong", i.e. non-corresponding, features in the image data.
Consequently, resulting image segmentations may be inaccurate in regions of high curvature.
This talk will discuss a method that overcomes these limitations. Its theoretical benefits are shown in synthetic experiments, while its practical applicability is evaluated in terms of segmentation accuracy and computational effort on exemplary highly curved anatomical structures.