| This dissertation presents methodologies and techniques to aid in the automation of the change detection process. The change detection process finds application in medical imaging specifically applied to lesion diagnosis and tumor detection. This study is limited to determining change in a previously selected window (i.e. local change), not change on a global scale. This is accomplished by the development of site model supported change detection algorithm. The change detection algorithm is divided into four main tasks: site model construction, preprocessing, registration, and change detection/quantification. Site model construction and preprocessing use classical signal and image processing techniques to derive the site model parameters (i.e. build the model). Registration, the most challenging component, uses a novel multi-step algorithm consisting of multi-object principle axis registration (PAR) for initial registration and thin-plate spline (TPS) transformation of control points for final registration. Three methods for combining the multiple transforms of initial registration are considered. They are local, average, and finite mixture. Local combination yields images containing discontinuity on boundaries. Average combination produces a smooth image, but assumes a rigid transform for the rest of the image. Finite mixture combinations produces a smooth image and can be used to model non-rigid deformation with several rigid transforms. In this study, finite mixture is used because the breast is generally assumed to be a non-rigid body. The change detection phase is performed by a two step process. Step one compares the joint relative entropy of the two image blocks with a detection threshold. Step two combines object area and center of gravity differences between the blocks as a means of quantification. |
