Smart manual landmarking of organs

Statistical shape models play a very important role in most modern medical segmentation frameworks. In this work we propose an extension to an existing approach for statistical shape model generation based on manual mesh deformation. Since the manual acquisition of ground truth segmentation data is a prerequisite for shape model creation, we developed a method that integrates a solution to the landmark correspondence problem in this particular step. This is done by coupling a user guided mesh adaptation for ground truth segmentation with a simultaneous real time optimization of the mesh in order to preserve point correspondences. First, a reference model with evenly distributed points is created that is taken as the basis of manual deformation. Afterwards the user adapts the model to the data set using a 3D Gaussian deformation of varying stiffness. The resulting meshes can be directly used for shape model construction. Furthermore, our approach allows the creation of shape models of arbitrary topology. We evaluate our method on CT data sets of the kidney and 4D MRI time series images of the cardiac left ventricle. A comparison with standard ICP-based and population-based optimization based correspondence algorithms showed better results both in terms of generalization capability and specificity for the model generated by our approach. The proposed method can therefore be used to considerably speed up and ease the process of shape model generation as well as remove potential error sources of landmark and correspondence optimization algorithms needed so far.