Shape Modelling: A Machine Learning Approach

Course Number 39525-01
Lecturers Marcel Lüthi, Thomas Vetter
Time and Location Wed 08:15 - 10:00; Seminarraum 00.003, Spiegelgasse 1
Wed 10:00 - 11:00; Computer-Labor U1.001, Spiegelgasse 1
Thu 09:00 - 10:00; Computer-Labor U1.001, Spiegelgasse 1
Start 18-02-2015
Prerequisites No specific formal requirements. Though students should have basic knowledge and skills in the fields of linear algebra, probability theory, statistics, and programming.
Objectives Upon successful completion of the course the students are able to:
- apply the theory of Gaussian processes for shape modeling
- understand how methods from machine learning can be used to learn properties of shapes
- implement a basic system for image analysis and shape modeling
- describe different methods for shape modeling
- describe different methods for image registration and see the connections between different classes of algorithms

Contents The course focus is on the problem of model-based image analysis. We discuss some basic approaches for image analysis and in particular focus on Analysis by synthesis. This approach is based on the assumption that in order to be able to analyse an image, one also needs to be able to synthesize this image by means of probabilistic models. Key questions of this course are how we can build such models, and use them for analysing (medical) images and the reconstruction of partial data.

We draw on the theory of Gaussian processes and study their application for modelling shapes. We take on a machine learning perspective, where the topic of modeling with Gaussian processes has been explored in depth and many algorithms and theoretical results have been established. By using those methods for shape modeling, the concepts can be visualized and thus it will be easier to obtain a good intuition of these methods. This aspect will in particular be covered in the exercises, where the theoretical methods are implemented to build a complete framework for model-based image analysis. The exercises will be done in the programming language Scala, but no prior knowledge of Scala is required.

The theory of Gaussian processes and Reproducing Kernel Hilbert spaces can be used to unify many distinct methods currently used in image analysis. We aim at making wherever possible the connections to popular research approaches in image analysis, shape modelling and registration, such that a student gets a good overview and understanding of the methods used in this field.
Literature Gaussian processes for machine learning, Carl Edward Rasmussen, 2006 (online version)
Pattern theory: from representation to inference, Ulf Grenander and Michel I Miller, 2007
Statistical Shape Analysis: Ian L. Dryden, Kanti V. Mardia, 1998
Assessment Lehrveranst.-begleitend

Please note: Student assessment will be based on a group project and a written report. Small student groups will implement a system for model-based segmentation of an anatomical structure. Organized as a competition between groups, the performance of each system will be evaluated using a publicly available benchmark dataset. A report, which summarizes the research findings in the project, will need to be handed in individually. The grading will be based on both, the group performance on the project and the scientific quality of the individual report.
Credit Points 4
Grades 1-6 0,5
Modules Vertiefungsmodul Bioinformatik (Bachelor Informatik 07)
Vertiefungsmodul Geoinformatik (Bachelor Informatik 07)
Modul Wahlbereich Informatik (BSF - Informatik)
Modul Praxis aktueller Informatikmethoden (MSF - Informatik)
Vertiefungsmodul Computational Intelligence (Bachelor Informatik 10)
Vertiefungsmodul Life Science-Informatik (Bachelor Informatik 10)
Registration Services (Requires login)