Research On The Fluid Registration Of Medical Images Based On SIFT

Non-rigid medical image registration plays an important role inclinical applications, such as monitoring treatment process, surgerynavigation, formulation of radiotherapy plan, retrospective studies andclinical training. In recent years, non-rigid registration based on physicalmodels has been paid more and more attention. In the fluid model, thedeformation of the target in images is regarded as the flow of fluid.Compared with other physical models, fluid model can be applied inlarger deformation. Traditional fluid registration algorithm take onlyintensity information as similarity measurement, which may cause falseregistration because of the noise and other interference in medical images.SIFT (Scale-invariant feature transform) method could extract local keypoints invariant to illumination change, scale change and rotation, whichcan provide stable matched point pairs for the registration. The shapeinformation, as the direct description of the target region, could alsoprovide stable matched point pairs. Because the shape of the target regionoften changes between preoperative and intraoperative images, a methodto extract deformation modes with PCA (Principal component analysis)method could be used by analyzing the deformation vectors of imagescollected in different treatment fractions, which could guide theamendment of the treatment plan and the following target contourextraction.So our research team proposed a fluid non-rigid registration methodcombining SIFT and shape information，which incorporates the correspondence of SIFT points and contour points into the construction ofsimilarity measurement. This paper is aimed to continue the study on2DSIFT algorithm and its extension to the3D space as well as the relatedalgorithms of deformation modes extraction method. All of the algorithmsare applied in prostate images to analyze their performance. First theSIFT method is used in prostate images in2D and3D space toquantitatively analyze the related parameters and the SIFT results arecompared between MRI and CT images. Then the experiment results ofdeformation modes extraction method verify the fact that the deformationspace could be characterized with a few limited deformation modes bycalculating the eigenvalues and eigenvectors of the deformation vectors.Three different methods to sample the deformation field of target regionare used to generate deformation vectors. Finally the fluid registrationmethod combining SIFT and shape information is applied to the prostateimages to experimentally determine the weight of image intensity, SIFTfeature points and the shape information in the construction of similaritymeasurement to obtain a better registration results.