Chief Market Risk Officer DnB NOR Markets, Oslo, Norway2010-Present
Practical Methodology for Inclusion of Modality-Specific Modifications in a Hierarchical Bayesian Deformation Model
An approach is presented which allows the incorporation of application-specific modifications to a general hierarchical shape deformation model. The general methodology models the perception of labeled points, or facets, across an image class through a joint distribution on facet position and image feature value. The modification introduces a set of parameters which represents the relative overall size of an image scene directly into the statistical model for the deformation. Through this, a more flexible and descriptive model is achieved without introducing an unmanageable computational burden. The methods are applied to and the modifications based on the application cardiac gated single photon emission computed tomography (SPECT). For this modality, a contraction factor and a center of contraction have real physical significance and are therefore included in the modeling. Results consistent with known heart behavior are seen for these quantities as well as for clinical quantities derived from the estimation results. A meaningful representation of the timeseries set of data is shown and improved results over traditional methods are seen. The method is also tested on two phantom datasets, one clinical and one mathematical, in order to quantify the ability of the method to track shapes and individual points, and good correspondence with known truth is seen. The methodology as described in detail is useful for any situation where automatic scaleability is useful. It also offers an instructive example of how the general hierarchical shape deformation model can be extended to incorporate application-specific information within the natural structure of the statistical model.