Finding and segmenting objects in medical images using a top-down strategy helps to cope with noise and artefacts in the data. Parametrising a model for specific detection or segmentation task can be tedious, however. Hence, we use object prototypes where object-specific variation is represented by a combination of general assumptions about variation and a task-specific decomposition of objects into sub-objects. The prototypes act as a top-down-realisation of a recognition-by-component strategy. Compared to a bottom-up recognition, the top-down strategy limits the generality of this approach but enables it recognition in very cluttered scenes. It also integrates segmentation of objects in noisy images into the analysis process.
Our model consists of a combination of deformable models. Objects deform under physical forces that are derived from image information. Model construction, communication between model parts and application to the detection and segmentation of objects will be discussed. Applications include the use of a prototypical model of shape and appearance to detect Heschl's gyrus in surface maps from MRI, the localisation of the substantia nigra in transcranial ultrasound images and the classification of ants from pictures of a database by comparing shape features.
back to the list of talks