Computational Anatomy and Registration
One of the most popular approaches employed for segmenting organ regions in given medical images is nonrigid registration of organ models to the images. There exist a huge number of model-to-image registration methods with several surveys of the registration methods in, e.g., [69, 120]. In this section, a very fundamental approach to registration, a MAP estimation, is described.
Many model registration methods first normalize the locations, shapes, and sizes of the patient bodies in given images (see also Fig. 2.6) and then register the organ models to the images. The normalization of the bodies in given images is required for decreasing the variation of the locations, sizes, and shapes of bodies of patients and for setting a common coordinate system to all of the training and test images. The organ models represent the statistics of the locations and/or of the shapes of target organs. In this section, two major statistical models of organs are explained: the probabilistic atlases and the SSMs.