Below you will find examples of research projects within the core research areas of the image group that you can use for inspiration in your application. On the basis of these short introductions, please elaborate a project description of length 3-5 pages (a maximum of 5 is enforced) and use the official application guidelines and application forms. You are welcome to contact potential supervisors if you have questions regarding your project description. For questions regarding the remainder of the application (CV, documentations of activities etc.) please read the guidelines or contact the PhD administration. As inspiration for your project description the project description of Sune H. Keller, who is currently doing his PhD in the The Image Group, is given here.
In this project we are developing image processing techniques to detect and quantify lung disease and its progression from CT scans for use in clinical trials. We are currently focusing on the quantitative analysis of lung emphysema -- a main component of chronic obstructive pulmonary disease (COPD, "smoker's lung", or "KOL" in Danish) and a major cause of death and disability worldwide. Accurate and reproducible measurement of the degree and type of emphysema is crucial to improve our understanding of the mechanisms involved in COPD, enable a better diagnosis and prognosis for individual patients, and facilitate monitoring of therapy in clinical trials.
In your project you may concentrate on one or several of the following topics:
On the topic of lung image analysis we have an active collaboration with Gentofte University Hospital in Copenhagen and AstraZeneca R&D in Lund, Sweden. They will provide the data and clinical expertise for this project.
I. Sluimer, A. Schilham, M. Prokop, and B. van Ginneken. Computer analysis of computed tomography scans of the lung: A survey. IEEE Transactions on Medical Imaging, 25(4):385--405, April 2006. [link]
Contact person and potential supervisor: Marleen de Bruijne.
This project aims at developing and evaluating machine learning and pattern recognition methods that are able to provide informative and interpretable descriptions of shape and/or appearance models of objects in image data.
Most often, in this type of modelling, the goal is to describe the whole class of shapes and/or appearances of interest simultaneously in an accurate and compact way. The typical unsupervised techniques employed for this purpose are (kernel) PCA, independent component analysis, and the like. However, generally, one is not interested in these shape and appearance models per se, but in some clinically or diagnostically relevant parameter. This implies that, when a clear task is involved, certain variations in shape and/or appearance are more important than others.
Employing statistical learning and information theory, methodologies are devised for investigating task-specific shape and appearance variations and determining the most informative ones. Besides enabling us to study such relevant shape and appearance changes, this research may also entail in improved modelling, e.g., for task-driven image segmentation problems.
Contact persons and potential supervisors: Marco Loog and Mads Nielsen.
Biological systems evolves under the influence of the physical properties of the surrounding world. For instance, it is believed that the visual system of humans and other animals have been shaped such as to be optimal under the physical conditions of the Earth. Humans perceive the environment by counting the photons hitting the retina producing electro-chemical impulses going to the visual centres of our brain. Based on these impulses the visual centres form hypotheses about the environment. The amount of information based on all the photon measurements is too enormous to actually be processed by the brain directly, hence the amount of information must be reduced by extracting relevant information from the photon counts.
This Ph.D. project focus on the development of a general phenomenological model of the extraction of relevant information in visual systems. Such a model should be task based, and we cast the problem of solving a task based on visual information as a probabilistic inference problem. The ensemble of physical environments give rise to what could be called the ecological statistics of the visual world. We can form models of this statistics in terms of stochastic image models. In order to be able to do probabilistic task based inference we also need a probabilistic description of the task of interest. Given these two models what are the most relevant features (revealing most information relevant for the task) we should extract from the image in order to make optimal inference for the task in question?
Martin Lillholm, Mads Nielsen, Lewis D. Griffin, Feature-Based Image Analysis, International Journal of Computer Vision, Volume 52, Issue 2 - 3, May 2003, Pages 73 - 95.
Pedersen, K.S. (2003). Statistics of Natural Image Geometry. Department of Computer Science University of Copenhagen.
Olshausen, B.A., & Field, D.J. (1996). Natural image statistics and efficient coding.Network-Computation in Neural Systems, 7 (2), 333-339.
Naftali Tishby and Fernando C. Pereira and William Bialek,"The information bottleneck method", Link, 2000.
Contact persons and potential supervisors: Kim Steenstrup Pedersen and Mads Nielsen.
The modelling of shapes often involves the detection/segmentation of a shape, the registration or correspondence between shapes, the description and representation of the space of shapes and finally doing useful statistics on this space. The modelling happens typically in a progressive approach but can also involve feedback and iteration for instance using the representation and the statistics upon it to improve the segmentation and so forth.
For elastic objects without the possibility of expert or automatic annotation of landmarks it is an open question how to summarize the shape appearance. Possibly implicit representations of surfaces could be registered with the simplest possible warp. The statistics could perhaps be expressed in terms of the likelihood of the necessary warp. Obviously this will - as usually in shape statistics - tie the registration and the statistics closely together.
This Ph.D. project will investigated the statistics of elastic shapes. A source of inspiration and primary application and evaluation area for the developed theory will be the modelling of Human Embryos. Optical 3D scans of Human Embryos with typical four blastomeres will be available as data. A possible question could be the statistical quantisation of intrinsic blastomere shape versus deformation caused by external interaction or statistics on ensembles of elastic objects.
T.F.Cootes, C.J.Taylor, D.H.Cooper, J.Graham Active Shape Models - Their Training and Application Computer Vision and Image Understanding Vol.61, No.1 January pp.38-59, 1995.
M.Nielsen, P.Johansen, A.D.Jackson, B.Laustrup Brownian Warps:A Least Committed Prior for Non-Rigid Registration Medical Image Computing and Computer-Assisted Intervention, 2002, Tokyo, Japan.
U.D.Pedersen, O.F.Olsen, N.H.Olsen A Multiphase Variational Level Set Approach for Modelling Human Embryos 2nd IEEE Workshop on Variational, Geometric and Level Set Methods in Computer Vision. In conjunction with ICCV, 2003, Nice, France.
Contact person and potential supervisor: Ole Fogh Olsen.