A Statistical Deformable Model for the Segmentation of Liver CT Volumes Using Extended Training Data

We present a fully automated method based on an evolutionary algorithm, a statistical shape model (SSM), and a deformable mesh to tackle the liver segmentation task of the MICCAI Grand Challenge workshop. To model the expected shape and appearance, the SSM is trained on 35 training datasets. Segmentation is started by a global search with the evolutionary algorithm, which provides the initial parameters for the SSM. Subsequently, a local search similar to the Active Shape method is used to refine the detected parameters. The resulting model is used to initialize the main component of our approach: a deformable mesh that strives for an equilibrium between internal and external forces. The internal forces describe the deviation of the mesh from the underlying SSM, while the external forces model the fit to the image data. To constrain the allowed deformation, we employ a graphbased optimal surface detection during calculation of the external forces. Applied to the ten test datasets of the workshop, our method delivers comparable results to the human second rater in six cases and scores an average of 68 points. Differences to the version presented at the workshop [1] are the increased number of training samples for the statistical model and some slight parameter optimizations.