Improved Landmark Initialization for 3D Statistical Shape Model Generation

Statistical shape models are used as a generic tool for medical image processing and image analysis. Their key advantage is the incorporation of prior knowledge about the shape variability of a certain anatomical object based on a set of representative training instances. The automatic identification of corresponding points (landmarks) across the entire set of training samples poses the essential challenge on the generation of the model. Minimization of the model's description length has proven to be an adequate strategy for this delicate matter. Thereby, a set of landmarks is distributed over each training shape, representing the initial set of corresponding points. These points are subsequently optimized in an iterative manner. The resulting landmark configuration however does not necessarily represent the shape of the structure of interest adequately. One possibility to overcome this problem is to employ a remeshing technique during and/or after the optimi-zation. Rather than interfering with the optimization method we chose a more appropriate initialization. This allows us to generate a homogeneous landmark distribution without em-ploying any remeshing steps. Moreover, we ensure that the algorithm converges to a global minimum since the landmark distribution adequately represents the object's shape at any time during optimization. In order to measure the goodness of the shape representa-tions we opt for a modification of the established landmark-distance based error measure, which takes surface similarity into account. Visual inspection as well as the computed quan-titative results show that the proposed initialization enables us to generate a model that outperforms the one resulting from employing the previously used initialization.