Using Points and Surfaces to Improve Voxel-Based Non-rigid Registration
MICCAI '02 Proceedings of the 5th International Conference on Medical Image Computing and Computer-Assisted Intervention-Part II
Computer Vision and Image Understanding
Parameterization-invariant shape statistics and probabilistic classification of anatomical surfaces
IPMI'11 Proceedings of the 22nd international conference on Information processing in medical imaging
| Hi-index | 0.00 |
Intensity-driven image registration does not always produce satisfactory pointwise correspondences in regions of low soft-tissue contrast characteristic of pelvic computed tomography (CT) imaging. Additional information such as manually segmented organ surfaces can be combined with intensity information to improve registration. However, this approach is sensitive to non-negligible surface segmentation errors (delineation errors) due to the relative poor soft-tissue contrast supported by CT. This paper presents an image registration algorithm that mitigates the impact of delineation errors by weighting each surface element by its segmentation uncertainty. This weighting ensures that portions of the surface that are specified accurately are used to guide the registration while portions of the surface that are uncertain are ignored. In our proof-of-principle validation, Monte Carlo simulations based on simple 3D phantoms demonstrate the strengths and weaknesses of the proposed method. These experiments show that registration performance can be improved using surface uncertainty in certain circumstances but not in others. Results are presented for situations when intensity only registration performs best, when intensity plus equally weighted surface registration performs best, and when intensity plus uncertainty weighted surface registration performs best. The algorithm has been applied to register CBCT and FBCT prostate images where the uncertainty of the prostate surface segmentation was estimated using contours drawn by five experts.