Premium
SU‐EE‐A3‐02: Incorporating a Priori Knowledge Into Deformable Registration Model
Author(s) -
Xie Y,
Xing L
Publication year - 2007
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 - 10.1118/1.2760359
Subject(s) - scale invariant feature transform , artificial intelligence , a priori and a posteriori , computer vision , image registration , computer science , feature (linguistics) , imaging phantom , robustness (evolution) , pattern recognition (psychology) , feature extraction , image (mathematics) , nuclear medicine , medicine , philosophy , linguistics , biochemistry , chemistry , epistemology , gene
Purpose: Conventional deformable registration treats every voxel in an image set equally. In reality, not all regions are equal: some parts are rigid and some are deformable. We investigate a strategy of using a priori knowledge of the system to improve the accuracy and robustness of deformable registration. Method: Our calculation consisted of two natural steps. First, the input images are auto‐classified into feature regions based on the intensity information. The feature regions in the fixed and moving images are matched as feature region pairs using SIFT (scale‐invariance feature transformation) method. Secondly, the established feature region pairs are used as pre‐determined control point associations to facilitate the thin plate spline (TPS) deformable registration. Because of the pre‐association of the feature pairs, there is no need to manually place the homologous control points, which is a difficult task (as detailed anatomy knowledge is often required) and has been a major source of inaccuracy. The proposed algorithm is evaluated by using a digital phantom and five sets of 4D CT images of different disease sites. Results: A method of incorporating prior knowledge into TPS deformable image registration has been developed. Using SIFT method, the feature region pairs can be easily identified. For example, for a region inside or close to a piece of bony structure, the correspondence can be found using SIFT because of the lack of deformation. In the digital phantom experiments, the new TPS algorithm is able to obtain a transformation matrix, which agrees with the known ones to within 3%. For the patients, similar level of success was achieved. A comparison with conventional TPS or BSpline approaches, the technique improves the accuracy and robustness. Conclusions: With incorporation of feature pairs, the deformable registration is significantly improved and much robust and accurate registration is attainable.