Propagating Interactive Segmentation of a Single 3-D Example to Similar Images: An Evaluation Study Using MR Images of the Prostate

In this paper we address the problem of segmentation propagation from one example onto similar images in an on-line and real-time fashion. As suggested in [1], we consider segmentation as a process consisting of two stages: the localization of the anatomy of interest and its boundary delineation. For each stage we identify and evaluate different potential candidate methods. All methods are applied on a dataset of 22 three dimensional (3-D) Magnetic Resonance (MR) images of the prostate. The high variation of appearance of the prostate across individuals is a challenging feature, which affects the repeatability of frameworks that leverage prior knowledge from one image example. Our observation is that the repeatability of the framework is improved, when a two stage processing strategy is realized, based on the deformable registration of [6,7], followed by Graph-Cuts [11,12] based segmentation.