Non-Rigid Registration Based Active Appearance Models for 3D Medical Image Segmentation

Abstract

Active shape models and active appearance models are getting increasingly popular in medical image segmentation applications. However, they are not suitable for three-dimensional (3D) images in their original form. This is due to the underlying shape representation (a point distribution model, PDM), which becomes impractical in 3D. Recently, it was shown that nonlinear registration algorithms can assist in the automatic creation of a 3D PDM. Based on this idea, we built a 3D active appearance model of brain structures. The model extracts the mean texture and the image deformation variation information from the training set of images. A special benefit is the inclusion of an extended region of interest into the model, making it suitable for segmentation of structures with poorly defined edges. We evaluated the model by applying it to the task of automatic segmentation of the hippocampi from magnetic resonance brain images. We found high accuracy of the model, which is comparable to the accuracy of the underlying registration method. The main benefit of the model-based segmentation over the registration-based segmentation is time, which is reduced from many hours (for registering an atlas to the image) to only a few minutes (for fitting the model to the image).