Joint Image Registration and Segmentation

Abstract

Image segmentation is ubiqitous in computer vision and image processing. In some applications such as in medical imaging, the problem may be very complex due to lack of sufficient image contrast, signal to noise ratio, volume avaraging, inhomogenities caused due to nonuniform magnetic field — in the case of MRI data sets, and sometimes lack of any real boundary due to the desired shape blending into the surrounding tissue. Such scenarios pose a formidable segmentation problem and provide motivation for incoporation of prior information in to the segmentation algorithms. There are several ways to incorporate priors into the segmentation formulation and we choose to use an atlas-based segmentation technique which requires that the given atlas shape/image be registered with the target image in order to find the desired shape segmentation in the target image thus requiring simultaneous registration and segmentation to be achieved. This is a natural choice for medical applications pertinent to some of the aforementioned scenarios. In this chapter, two approaches are presented, one is a pure PDE-based formulation and the other is a vriational formulation involving a level-set representation of the shape being segmented. The former approach copes with 3D non-rigid motion and the later although easily applicable to the non-rigid motion case is shown here only for the 2D rigid motion case. Efficient numerical methods are used in both the approaches and the performance of the algorithms is depicted via examples on synthetic and real 3D and 2D image data.