Fast non-rigid registration and model-based segmentation of 3D images using mutual information

In this paper, we present a method to register medical images non-rigidly and to determine critical structures using a model-based segmentation. We present our method in the case of 2D and 3D magnetic resonance images (MRI) of the brain. In our approach, we first use an existing algorithm for rigid matching of medical images by Mutual Information maximization for the initialization of our registration. Then we apply our gray level based non-rigid matching algorithm to match the contours of the model on acquired medical images. We have also added an elastic internal energy term to our algorithm, so that the contour deformations are elastically propagated throughout the whole object to be matched. We present this non-rigid and iterative method in the context of inter-patient matching and of deforming an elastic brain atlas on medical images for automatic localization and identification of brain structures in medical images. The main advantages of our implementation are its very low computational cost in comparison to other methods and the fact that it can be applied directly on a synthetic atlas without passing through a corresponding medical image. We present preliminary results, we mention some existing limitations of our method and indicate future work.