For coregistration of medical images, rigid methods often fail toprovide enough freedom, while reliable elastic methods areavailable clinically for special applications only. The number ofdegrees of freedom of elastic models must be reduced for use inthe clinical setting to archive a reliable result. We propose a novel geometry-based method of nonrigid 3Dmedical image registration and fusion. The proposed method uses a 3D surface-based deformable model asguidance. In our twofold approach, the deformable mesh from oneof the images is first applied to the boundary of the object to beregistered. Thereafter, the non-rigid volume deformation vectorfield needed for registration and fusion inside of the region ofinterest (ROI) described by the active surface is inferred fromthe displacement of the surface mesh points. The method was validated using clinical images of a quasirigidorgan (kidney) and of an elastic organ (liver). Thereduction in standard deviation of the image intensity differencebetween reference image and model was used as a measure ofperformance. Landmarks placed at vessel bifurcations in the liverwere used as a gold standard for evaluating registration resultsfor the elastic liver. Our registration method was compared withaffine registration using mutual information applied to thequasi-rigid kidney. The new method achieved 15.11% better quality with ahigh confidence level of 99% for rigid registration. However,when applied to the quasi-elastic liver, the method hasan averaged landmark dislocation of 4.32 mm. In contrast, affineregistration of extracted livers yields a significantly (P=0.000001) smaller dislocation of 3.26 mm. In conclusion, ourvalidation shows that the novel approach is applicable in caseswhere internal deformation is not crucial, but it has limitations incases where internal displacement must also be taken into account.
展开▼
