Erasmus MC : ASCI

Onderzoeksgroepen van de Technische Universiteit Delft, Vrije Universiteit, Universiteit van Amsterdam, Universiteit Leiden, Universiteit van Utrecht, Universiteit Twente, Universiteit Groningen, Technische Universiteit Eindhoven, Erasmus Universiteit Rotterdam en de Radboud Universiteit Nijmegen participeren in ASCI.

De Vakgroep Medische Informatica participeert in ASCI in een aantal cursussen.

a23 Bioinformatica

Doel

Studenten die deze cursus hebben gevolgd kunnen wetenschappelijke literatuur op het gebied van data-driven bioinformatics lezen, begrijpen en bediscussieren. Waneer een aantal eenvoudige implementaties van de algoritmes genoemd in de literatuur wordt gegeven, dan kunnen studenten daarmee zelf een geintegreerde data-driven bioinformatica toepassing maken.

a22 Knowledge driven Image Segmentation

Doel 

Het geven van een goed inzicht in de laatste ontwikkelingen op hetb gebied van knowledge-driven segmentatie methoden.

Cursus inhoud 

Deze informatie is nog niet in het Nederlands vertaald.

The course will address three main topics:
1. Deformable models
Lecturer: dr W.J. Niessen (dept. Medical Informatics Erasmus MC)
Deformable models have become a popular tool for (semi-) automated (medical) image segmentation. A deformable model is a contour or surface, which deforms in order to capture objects to be segmented. The deformation is guided by forces, which can be determined by features in the image (edges, texture) to be segmented, but also by geometric constraints (smoothness of the curve or surface; prior information of the shape to be segmented). The trade-off between geometric and image-derived information lies at the basis of the popularity and diversity of deformable model based methods; if image information alone is insufficient for a proper segmentation, the combination with geometric constraints may still yield plausible solutions. In this part of the course we consider two classes of deformable models:
Explicitly parameterized models and implicitly parameterized models. In the first class of models, often referred to as snakes, a parameterized curve or surface is evolved in time to capture object boundaries.  In the second class of models, often referred to as level sets, the curve or surface is represented by the zero level set of a higher-dimensional function. This representation has certain advantages, most notably that a change in topology of the curve or surface during the evolution is effectively facilitated. For both classes issues related to initialization, optimization, user interaction, and the incorporation of prior knowledge will be discussed. Furthermore, a large number of applications will be shown.

2.Statistical models
Lecturer: dr ir B.P.F. Lelieveldt
Statistical models capture the shape of an object from a training population as an average shape and a number of characteristic variations. These models have been initially developed for shape analysis and gaining insight into typically occurring anatomical variations. Apart from shape analysis, the trained eigenvariations can also be applied to image segmentation by restricting the search space for the model matching to statistically plausible directions. Contrary to the deformable models mentioned earlier, they integrate population based a-priori knowledge about shape and image appearance into the segmentation, further increasing robustness. During this course, two types of statistical models will be treated. Active shape models describe the distribution of a set of characteristic landmark points. ASMs can be used for image segmentation by using local intensity models to find update points for each landmark, and deforming the model within the statistically trained limits. Active Appearance Models simultaneously describe the shape and the intensity of the object of interest as seen in an image patch. Like Active Shape Models, a model of the landmark distribution captures the shape variability, whereas intensities are modeled by mapping the object patches to the shape average, and determining gray value eigenvariations. Model matching is realized by deforming the model in such a way, that it "blends in" with the target image, again constraining the deformation to the statistically trained limits. 
The course will treat the basic concepts behind Active Shape and Appearance models (2D and 3D), and will discuss several medical and non-medical applications (facial recognition, object tracking). In addition, a more superficial overview of alternative statistical modeling methods, such as medial models, statistical deformation models and probabilistic atlases is provided.

3. Pattern recognition approaches to image segmentation
Lecturer: Dr. M. de Bruijne (IT University Copenhagen and dept. Medical Informatics Erasmus MC)
Image segmentation can be formulated as a pattern recognition problem. Once an image is divided into elements, one can segment an object of interest by classifying each element as belonging to the object or not. To this end, features need to be computed for each element, and a classifier must be trained to map feature vectors into object labels. The elements can be simply voxels or pixels, or other primitives, such as the output of a watershed segmentation or line elements obtained by ridge detection etc. The basic concepts of such segmentation schemes will be presented and illustrated with a range of applications. Several techniques for post-processing classification results will be discussed, such as iterative relabeling, relaxation labeling or Markov random field models. In addition, it will be demonstrated how active shape and appearance model segmentation schemes can also benefit from the use of more complex features and classifiers that are not used in their original formulation.