Scale Space and Medical ImagingProf. Bart M. Ter Haar RomenyImage Sciences Institute, University Hospital Utrecht, <E01.334, > The Netherlands. Course Schedule :

Abstract :
The course focuses on multiscale image analysis, its relation to human vision,
and the differential structure of images in a modern,
physics based approach.
Image analysis is the extraction of useful information from images. It has become clear that a multiresolution approach is indispensable. The physical reason is the fact that we deal with an observed signal, where the aperture size of the measurement device can be varied over some range, the scalerange. It turns out that the human visual system also widely exploits a diversity of multiscale filters in its processing layers.
In this course a modern mathematical (and physics based) approach to multiscale image analysis is presented as a branch of computer vision. We give an intuitive introduction to multiscale image analysis, trying to keep the analogy with stages in the human visual system as close as possible.
Among the topics covered are: the physical basis of regularization, the design of differential invariant features in images to high order, multiscale analysis of 2D and 3D shape, multiscale motion analysis from image sequences, depth from stereo, orientation analysis, and the use of contemporary, wellunderstood mathematical tools in this field such as differential geometry and tensor analysis.
The majority of the examples discussed are from 2D, 3D and 4D (3Dtime) medical imaging. We devote some time to the efficient numerical implementation of the different techniques. Experiments in Mathematica 3.0 illustrate the theory and applications in practice.
Topics :
Literature :
A set of readers with selected papers and copies of the sheets will be handed out during the course.
Reading suggestions:
Example of an application of scalespace theory to medical imaging: 3D volume
visualization of a 7 week old human phoetus in the belly. Left: original data. Right:
3D nonlinear diffusion technique to reduce the noise while preserving edges. The
implementation in 3D uses ultrafast numerical methods. Technique: Joachim Weickert,
PhD, Utrecht University, 1997. Image acquisition: Kretz Combison 500, 256^{3}
cubic dataset, interpolated from the original conebeam echo acquisition.