We recently introduced labeling of discrete Markov random fields (MRFs) as an attractive approach for non-rigid image registration. Our MRF framework makes use of recent advances in discrete optimization, is efficient in terms of computation time, and provides great flexibility. Any similarity measure can be encoded right away, since no differentiation is needed. In this work, we investigate the performance of our framework in a challenging scenario: the registration of thoracic CT images. In order to assess the potential of the discrete MRF setting, we employ the simplest registration objective function based on intensity differences. The registration is fully-automatic, constant parameters are used throughout the experiments, we omit the use of the available segmentations, and (except for linear pre-alignment) no pre-processing of the data is performed. Despite the simplicity of our experimental setup, we are able to obtain accurate registration results for most of the data in a very efficient manner. Our registration software is freely available http://www.mrf-registration.net/.     «

We recently introduced labeling of discrete Markov random fields (MRFs) as an attractive approach for non-rigid image registration. Our MRF framework makes use of recent advances in discrete optimization, is efficient in terms of computation time, and provides great flexibility. Any similarity measure can be encoded right away, since no differentiation is needed. In this work, we investigate the performance of our framework in a challenging scenario: the registration of thoracic CT images. In or...     »