Medical images registration with a hierarchical atlas

Atlas-based medical image segmentation is a well known method for including prior knowledge in medical image analysis. It requires as basic component the registration of an atlas with the image. In this paper, we introduce the concept of hierarchical atlas and show how to efficiently include it in a state-of-the-art non-rigid registration algorithm. We first present how to build a hierarchical atlas. Then we present the extension of a non-rigid registration algorithm, namely the B-spline Free Form Deformation (FFD), to a hierarchical version. The procedure includes first an affine registration to bring the atlas and the patient image in global correspondence. Then the non-rigid registration is performed layer by layer, i.e. registering the image with each layer of the hierarchical atlas, using the result of the registration of the previous layer as initial condition for the registration of the next layer. We show on 2D CT images that this approach gives better results than the non-rigid registration algorithm alone, in terms of registration accuracy.