Initialization of deformable models in 3D magnetic resonance images guided by automatically detected phase congruency point landmarks

Deformable models are a widely used approach for 3D medical image segmentation, due to its flexibility and capability to incorporate prior anatomical knowledge in the segmentation process. However, methods based on deformable models are, usually, very sensitive to initialization, requiring that the initial position and shape of the model are as close as possible to the structure of interest in the target image. Thus, we propose in this work a novel approach for automatic initialization of deformable models for 3D MR images, using a set of automatically detected point landmarks to guide the process. Our approach combines 3D phase congruency based landmark detection, shape context based descriptors, nearest neighbor search and multilevel non-rigid B-spline transform estimation. A freely available atlas of 3D triangular meshes of brain structures, aligned to a reference image, is used as source for models. Our approach was tested in the initialization of models representing the corpus callosum (CC), left hippocampus (LH) and right hippocampus (RH). Results have shown a significant increase in the Jaccard and Dice metrics after the models were initialized by our method.