Statistics Seminars: Bayesian registration and shape analysis of object data, with applications to proteomics and medical imaging
3 December 2012 14:00 in CM221
We consider the Bayesian analysis of object data, such as functions, images and shapes. Of fundamental interest in comparing object data is the separation of registration information (e.g. translation, rotation, scale and reparameterization) and shape information (what remains). However, there is inherent non-identifiability in separating these sources of information. The user must provide some prior beliefs about the registration information, and so a Bayesian approach is very natural. The separation of registration and shape information is endemic in a wide range of applications, including in bioinformatics and neuroscience.
A key issue in dealing with registration is whether to choose a model in the ambient space of the objects, or in the quotient space of objects modulo registration transformations. In the former case the distributions can become complicated after integrating out the registration information, whereas in the latter case the geometry of the space can be complicated. Although in general these approaches are different, in many applications there are often practical similarities in the resulting Bayesian inference due to a Laplace approximation.
We consider several applications to illustrate these issues, including registering mass spectra and molecules in proteomics, and comparing 3D curves and surfaces in medical imaging.
Contact firstname.lastname@example.org for more information