Premium
Automated segmentation of the prostate in 3D MR images using a probabilistic atlas and a spatially constrained deformable model
Author(s) -
Martin Sébastien,
Troccaz Jocelyne,
Daanen Vincent
Publication year - 2010
Publication title -
medical physics
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 1.473
H-Index - 180
eISSN - 2473-4209
pISSN - 0094-2405
DOI - 10.1118/1.3315367
Subject(s) - segmentation , probabilistic logic , artificial intelligence , atlas (anatomy) , computer science , computer vision , image segmentation , robustness (evolution) , sørensen–dice coefficient , statistical model , pattern recognition (psychology) , scale space segmentation , image registration , image (mathematics) , biochemistry , chemistry , gene , paleontology , biology
Purpose: The authors present a fully automatic algorithm for the segmentation of the prostate in three‐dimensional magnetic resonance (MR) images. Methods: The approach requires the use of an anatomical atlas which is built by computing transformation fields mapping a set of manually segmented images to a common reference. These transformation fields are then applied to the manually segmented structures of the training set in order to get a probabilistic map on the atlas. The segmentation is then realized through a two stage procedure. In the first stage, the processed image is registered to the probabilistic atlas. Subsequently, a probabilistic segmentation is obtained by mapping the probabilistic map of the atlas to the patient's anatomy. In the second stage, a deformable surface evolves toward the prostate boundaries by merging information coming from the probabilistic segmentation, an image feature model and a statistical shape model . During the evolution of the surface, the probabilistic segmentation allows the introduction of a spatial constraint that prevents the deformable surface from leaking in an unlikely configuration. Results: The proposed method is evaluated on 36 exams that were manually segmented by a single expert. A median Dice similarity coefficient of 0.86 and an average surface error of 2.41 mm are achieved. Conclusions: By merging prior knowledge, the presented method achieves a robust and completely automatic segmentation of the prostate in MR images. Results show that the use of a spatial constraint is useful to increase the robustness of the deformable model comparatively to a deformable surface that is only driven by an image appearance model.