Intrinsic Random Fields and Image Deformations

The stochastic structure of images, especially individual medical images as they are reconstructed nowadays from arrays of medical imaging sensors, is becoming steadily better understood. Less attention has been paid to the parallel notion of estimation error for the deformations that convey relations among these images, such as localized abnormality or growth prediction. The dominant current formalism for the biostatistics of deformations deals solely with the shape of a set of landmarks parameterizing the deformation, not otherwise with its behaviour inbetween the landmarks. This paper attempts to fit a rigorous stochastic model for a deformation between landmarks and to assess the error of the fitted deformation. The relation between two images is modelled as a stochastic deformation, i.e. as an identity map plus a stochastic process whose value at every point is a vector-valued displacement. There are two common strategies for fitting deformations given information at a set of landmarks. One involves minimizing a roughness penalty, e.g. for a thin-plate spline, and another involves prediction for a stochastic process, e.g. for a self-similar intrinsic random field. The stochastic approach allows parameter estimation and confidence limits for the predicted deformation. An application is presented from a study of breast images and how they deform as a function of the imaging procedure.