1 page
poster in PDF format.
December 2001
Medical Image Display and Analysis Group (MIDAG)
http://midag.cs.unc.edu/
Jointly organized by the Divisions of Applied Mathematics and Engineering, and the SHAPE Lab., Brown University.
1 page
poster in PDF format.
Room 190, B&H (Engineering Bldg.)
Monday Dec.3, 2 hours incl. questions, and Tuesday Dec.4, 2 hours.
The geometry of anatomic objects can be described using a variety of different atoms: voxels, landmarks, boundary locations with or without orientation, and medial descriptors. This geometry can be used as the basis for model-based object segmentation from a (typically 3D) image, for image registration, or for characterization of an object's shape. Probability distributions for geometry or shape can be used as priors in posterior optimization methods for segmentation or registration, or they can be used to discriminate classes of shapes, such as normal and pathological variants of an object. In this tutorial I will describe the various atoms and the sampled and parametrized object descriptions that can be built from them, as well as the spatial correspondences across deformation to which they lead, and I will compare these descriptions. Issues of the scales of representation provided and the need for multiple scales will be faced. The means of representing and training probability distributions of geometry will be faced. Metrics for geometric typicality based on pure geometry and metrics based on geometry and probability will be described.
Room 190, B&H (Engineering Bldg.)
Tuesday Dec.4, 4pm-5pm.
3D segmentation of objects in 3D images can be usefully done by deforming structural models of anatomic objects into target images, the models having been built from training images. M-reps, sampled medial volume representations, are a means of representing the models that are particularly apt in this framework because they have special capabilities in deformability and intuitiveness, and provide good capabilities of many levels of scale and thus efficiency for any level of performance. After briefly describing m-reps models, this talk will focus on multiscale methods for deforming them into target images to segment objects of interest for radiotherapy treatment planning of the abdomen and male pelvis and for neuroscience of brain structures. These methods involve successive optimizations of objective functions summing a geometric typicality measure and a geometry to image match measure. Each of these measures will be described, as will the role, in each, of the spatial correspondence provided by m-reps. Results of segmenting a number of anatomic structures from CT and MR images will be illustrated. Quantitative results comparing inter-human kidney segmentation differences to mreps-to-human segmentation differences will be given.
Poster
(from MICCAI'01) as a PDF file.
[Co-authors: James Damon, Thomas Fletcher, Paul Yushkevich]
Room 110, Div. of Applied Maths.
Wednesday Dec.5, 4pm-5pm.
As has not been previously appreciated, the geometry of the transformation from boundaries to medial loci and the geometry of the transformation from medial loci to boundaries is not entirely the same. This fact leads to a number of interesting options for the discrete representation in 3D of medial loci and for interpolating the continuous medial locus and the continuous boundary corresponding to the discrete medial representation. Moreover, it is advantageous to look at medial geometry not simply in terms of the zeroth order medial position (x) and medial width (r) functions but also in terms of the medial frame and the medial object angle, which are first order in x and r. With this zeroth plus first order medial geometry, the medial locus can be usefully described as a manifold of medial atoms, each consisting of a medial position and two length r vectors with tails at the medial position x that we call medial sails, each having a head incident to and orthogonal to the corresponding boundary position. This view induces a useful "figural" correspondence between 3-space locations relative to deformed versions of single-figure or multifigure objects, and we hypothesize that an object-relative non-Euclidean metric space can be built upon this scheme.
Local organization: Prof. David Mumford.
Last Updated: Dec. 5, 2001
Webmaster: F.Leymarie