|Stochastic Systems Group|
SSG Doctoral Student
Segmentation of the prostate in MR images is a difficult task for trained radiologists, let alone for trained computers. Using traditional segmentation techniques (e.g., statistical classification, curve evolution, etc.) is not very useful due to the overwhelming number of different tissue types in the pelvic region as well as weakly defined edges for the prostate itself. Other difficulties arise from using an endorectal surface coil which enhances SNR, but creates a large multiplicative bias field.
We exploit the signed-distance map shape models of Leventon et al to impose a constraint on the space of admissible solutions. Shape priors are constructed by decomposing manually segmented training data into eigenshapes using PCA. Final segmentations are obtained by minimizing an appropriate energy functional over the space of parameters defined by the eigenshapes and pose estimation. We discuss limitations of the Leventon shape models, the difficulties of establishing a ground truth for the correct segmentation, and current research avenues.
Problems with this site should be emailed to email@example.com