Topological active volumes for 3D image segmentation

  1. Barreira, Noelia
Supervised by:
  1. Manuel Francisco González Penedo Director

Defence university: Universidade da Coruña

Fecha de defensa: 16 June 2009

Committee:
  1. Petia Radeva Chair
  2. Antonio Blanco Ferro Secretary
  3. Xosé Manuel Pardo López Committee member
  4. Jordi Vitrià Marca Committee member
  5. Aurelio Joaquim de Castro Campilho Committee member

Type: Thesis

Teseo: 226501 DIALNET

Abstract

The recent development of 3D acquisition technologies has emphasised the need of algorithms for the understanding of volumetric datasets. In this sense, object segmentation and reconstruction are important tasks in the image processing pipeline. The segmentation task isolates the object points from the background in order to create a coherent and consistent model of the detected structures. Unfortunately, the detection of objects in volumetric datasets is not straightforward due to the complex topology of the objects and the computational cost of working in a 3D space. This PhD thesis proposes a new deformable model, the Topological Active Volumes, focused on segmentation and reconstruction of volumetric images. Deformable models are widely used in image analysis and provide a general framework that can be applied to solve different problems in specific domains. They are based on a contour, surface or volume that evolves under the influence of energy functionals related to the contour, surface or volume and the input image. The energy functionals ensure the model continuity and smoothness and are defined in such a way that their minimum coincides with the features of interest. Thereby, the evolution process is just a minimisation process. The Topological Active Volumes integrate features of region based and boundary based segmentation methods in order to perform a volumetric segmentation and, this way, fit the contours of the objects and model their inner topology. Also, this model implements automatic procedures, the so called topological changes, that alter the mesh structure in order to segment complex features, such as pronounced curvatures or holes, and detect several objects in the scene. In addition, this work proposes two segmentation methodologies based on minimisation techniques with different scopes. The suitability and accuracy of the proposed model and segmentation methodologies are proven in both synthetic and real 3D images with different levels of complexity.