New Insights On Multidimensional Image And Tensor Field Segmentation

New Insights on Multidimensional Image and Tensor Field Segmentation

Extracting knowledge from images through feature extraction is a topic of paramount importance for the Image Processing and Computer Vision communities. Within this general objective, this thesis focuses on the combination of the intensity and texture information, encoded by means of the local structure tensor (LST), for the segmentation of images. The LST is a well-stablished tool for the representation of oriented textures, and its incorporation to the segmentation process has reported to improve the segmentation performance. However, its combined use with the intensity is a complex issue that must be tackled carefully. This dissertation explores various alternatives to achieve this combination, and besides studies the problem of the balance of both sources of information. Within a level set framework, the segmentation is first performed in the tensor domain based on the definition of novel LST tensor variants that incorporate intensity information. A different approach is also considered based on a common energy minimization framework that allows the usage of both the insensity and the LST respecting their most adequate representation forms and suitable metrics. Besides, an adaptive procedure for the determination of the weighting parameters is proposed that takes into account the respective discriminant power of both features. The segmentation of tensor fields is also addressed in this dissertation. In this direction, an extension to the state-of-the-art approaches for the segmentation of tensor data has been derived which is based on the modeling of tensor data using mixtures of Gaussians. The application of this scheme can be devoted to the combined use of the intensity and texture as introduced before, as well as for the stand-alone segmentation of tensor fields. The methods proposed in this dissertation are applied to three medical image applications. The first two are performed using both the intensity and the LST in a combined approach as proposed in this thesis. Specifically, the segmentation of hand bones from radiographs is first addressed, related to the problem of the automated determination of the skeletal age in children. Next, the endocardium of the left ventricle is extractred from 3D+T cardiac MRI images. The third application is devoted to the segmentation of the corpus callosum from diffusion tensor MRI, and is thus an application of the Gaussian mixtures model for tensor field segmentation.

New insights on multidimensional image and tensor field segmentation

New Methods to Improve Large-Scale Microscopy Image Analysis with Prior Knowledge and Uncertainty

Multidimensional imaging techniques provide powerful ways to examine various kinds of scientific questions. The routinely produced data sets in the terabyte-range, however, can hardly be analyzed manually and require an extensive use of automated image analysis. The present work introduces a new concept for the estimation and propagation of uncertainty involved in image analysis operators and new segmentation algorithms that are suitable for terabyte-scale analyses of 3D+t microscopy images.