Speaker: Frederik Maes, UZ Leuven, Medical Imaging Research Centre
UCL Contact: Dominique Drai (Visitors from outside UCL please email in advance).
Date/Time: 21 Jun 12, 12:00 - 13:00
light lunch/refreshments will be provided.
Medical image computing relies on models to cope with the complexity of the anatomical scene and of the imaging data itself. A fundamental problem in medical image computing is the modeling of anatomical shape and shape variability. Various shape models have been developed that make use of an explicit representation of the object's boundary. Alternatively, shape can also be represented implicitly, namely in the form of an image, such as an atlas. In this case, shape matching boils down to image registration. Image registration has long been a challenge by itself in medical image analysis, for instance using intensity based similarity measures such as mutual information. Traditionally, atlas-to-image registration and atlas-guided image segmentation would be considered as separate problems. However, by considering both problems simultaneously, the registration can benefit from intermediate segmentation results, such that the similarity measures driving the registration process are based on object labels directly instead of their intensities. By extending the segmentation of a single image to the joint segmentation of multiple images simultaneously, the use of image registration has evolved into a computational strategy for image segmentation, atlas construction and image based clustering. These novel schemes offer the possibility of unsupervised group-wise analysis to extract the typical morphological patterns that discriminate between different subgroups in the population. The challenge now is how to apply these models to the analysis of new images of individual patients.