Multi-modality Medical Images Registration Based On Maximization Of Mutual Information With Edge Information

It is common for patients to undergo multiple tomographic radiological imagingfor the purpose of medical diagnosis, treatment planning and navigation in surgicaloperation. These three dimensional modalities provide complementary informationabout pathology, and associated normal anatomy. But variations in patient orientation,and differences in resolution and contrast of the modalities make it difficult for aclmician to mentally fuse all the image information accurately. For this reason, therehas been considerable interest in using image registration techniques to transfer allthe information into a common coordinate frame.Medical image registration is defined as a one-to-one mapping between the coordinates in one space and those in another such that all corresponding points in thetwo images have the same anatomical structure.Maximization of mutual information ( MI ) developed by Col-lignon, Wells andStudhohne, etc. has been widely used in clinical practice since 1995, because of its superiority over other registration algorithms in accuracy and robustness.Mutual information is a basic concept in information theory, which is usually used to measure the statistical dependence between two random variables, or the amount of informatiqn that one variable contains about the other. The method presented in this paper applies mutual information to measure the information redundancy between the intensities of corresponding voxels in both images, which is assumed to be maximal if the images are geometrically aligned.The mutual information of two images is evaluated :where P_FR(f,r) is the jiont image intensity distribution,P_F(f) and P_R(r) are the marginal image intensity distributions of floating image F and reference image R respectively.Based on the maxiiniation of mutual information method need not make some other priori suppose to multi- modal medical images, do not need to carry on segment and pre-processing to the image, and the accuracy can reach the inferior rank of pixel.But the method of based on the maxiiniation of mutual information exists robust problem,what is said,in the moment of searching optimical transformation exists the disterbance of great numeber of local ex-tremums in order to registration running into the local extremum to lead to misregistration. No matter what reason leads to robust problem,its essential question lies in the mutual information defend on the pixels without enough correlative information.So the mutual information is instability and leada to misregistration. Based on this reasons,my paper puts forward grey level change violent high-frequency information whether border information of image is it arrives original mutual information.This can increase the relativity between images and thereby improve robust.The mutual Information with edge information of two images is evaluated :I*(XY, a) = H*(X, a) + H*(Y, a) - H*{XY, a)where H*(XY,a) is the jiont entropy,H*(X, a) and H*(Y,a) are the Shannon entropy with edge information of floating image X and reference image Y respectively.From the relation of entropy and mutual information,we know that if want to improve the mutual information model.it is a straight approach from Shannon entropy.Fisrt,we compute the edge probability of images. We define N is the pixels pairs summation .of def-ferent intensity in the images,n,j is a some intensity i the pixels pairs summation with its defferent intensity conterminous pixel , so have :as well as 2D image is qua example,n,j we use " four conterminous points " to computers following image :\/\ \////\sis one of edge pixel points with the intensity i , p/,-, k = 1, 2, ? ? ? ,4 are four its conterminous points,we want to compute how much defferent pixel pairs ( s , pk ) have. We can compute N to all the edge points of the image to use this method.At the time , we use pi(i) to show the edge probability with intensity i in the image,so have :v—^TinNthis one is our the imgae edge information.thereout,we can improve the Shannon entropy H(X) = — \^ P{xi)l°9P{iwe say thatH*(X) = -is the Shannon entropy with edge information.Considering that the coordinates may not be integer values after the spatialtransformations, we introduce several interpolation techniques into voiding the non-integerintensities and detailedly show the PV interpolation technique, which would make trouble for computing the histograms. Because the computing of the MI relies on the intensity values, the PV interpolation is our perfectchoice, which does not produce fractional or new intensity values.So this technique is widely used.But in the PV interpolation computes the weighted only to depend on four near points around the interpolation one, we think that this is not enough compellent.In fact,Ta(s)'s grey intensity is not only relative with four near points around the interpolation one, but also relevant with the probability detensity of the four points respectively.Therefor their combination will be suitably.We bring forward the improved PV interpolation arithmetic : n. dxThe improved PVsimilarly,we obtain joint intensity histogram to use four points around TQ(s).The weight u* in the improved PV interpolation arithmeticdistributes as the following :i=l 4I Ithereinto,a;i = dx ■ dy,u)2 = dx ? (1 — dy),LJ3 — (1 —(1 — dx) ? (1 — dy), this is coincident with the PV arithmetic,^.,;4 4I,---,4.Through the results of the experiments, we come to the conclusion that the maximal MI with edge information registration method is asubpixel-level accurate, highly robust, completely automatic registration method of multimodality medical images. After speeding up the algorithm, it is perfect to apply to the clinic.The whole framework of this thesis is listed as follows:one is the prolegomenon of this paper;chapter three chapter three is the sim-plereviews of the medical images registration techniques;introduces the theory of entropy and then the theory of the MI; chapter four describes the MI criterion of image registration and my paper's mainly...