PDF version of this document
The task of image analysis, especially in the bio-medical domain, must take into consideration the variation in shape and appearance of objects. The invariant presumption is that corresponding objects in all images are of one particular class so we can typify the contents of the image by training an entity that captures inter-subject variation as well as atrophies.
Statistical analysis of shapes [3] which obtains a model of deformation goes back a decade ago. The principles were later extended to sample the variation in pixel intensities (also commonly referred to as textures) to create a model of full variation that is able to synthesise full appearances [4] and their successful application to medical data has been frequently demonstrated [5]. The correlations between shape and intensity are learned using Principal Component Analysis [6] where much of the power of these principles lies.
The integrity of models breaks down if correspondences, annotated in the form of spatial landmarks, are inappropriately identified. Furthermore, the annotation process involves a preliminary segmentation process which highlights parts of the data where landmarks can and should be placed. Although this has become a solved problem in statistical modeling of shape, it is yet difficult to select good landmarks in images which strive to retain full appearances rather than contours or surfaces solely. Several attempts have been made to resolve the issue [7,8,9], but none was optimal or even quite satisfactory. Alignment has become the means by which this crucial limitation can be solved and the foundations of image registration assist in establishing this alignment.