Multimodal Medical Image Registration Based On Mutual Information

Medical image registration which was developed from 1990s is one of importantresearch area of medical image processing. It is widely applied in clinic. Mutualinformation based registration is one of important method, but it ignores many spatialinformation of image. In this thesis, the principium and drawback of maximizationmutual information registration method are analyzed. Fuzzy mutual information, anew registration measure which extends from mutual information, has beenestablished. The thesis continues with the experiments of availability and robustness.A 3-D medical image registration experiment system based on mutual information isdesigned and implemented.The main work and contributions of the dissertation include:(1) Starting with the matrix description of digital image, the joint intensityordered pair matrix and intensity ordered pair matrix of are given. The mathematicalmatrix description of entropy and mutual information are presented. Thecharacteristics and drawbacks of maximization mutual information registrationmethod are analyzed.(2) The intensity correlation phenomenon of images is described, and physicsbasis and fuzzy characteristics of the intensity correlation phenomenon are alsoanalyzed. The concept of intensity correlation function of registration is developedfrom the analysis.(3) The mathematical description of fuzzy mutual information is based on theintensity correlation function of registration, so the fuzzy mutual informationregistration measure is established. The effects of image size and grey number to thefuzzy mutual information are analyzed, and the fuzzy mutual information registrationmeasure is normalized in order to decrease the effects of overlap region change tofuzzy mutual information.(4) Detailed experiments on the multimodal medical image database ofVanderbilt University demonstrate that the fuzzy mutual information measure hasmuch better availability and robustness for different interpolation methods, partialoverlap and image noise than mutual information measure.(5) The background of medical image is classified as single grey background andnoisy background according to the grey number of background. A background noisy processing algorithm which decease the affects of background to mutual informationis presented. The local extremum of mutual information can be enhanced when imagebackground has been processed through binarization, erosion, dilation.(6) Implementation methods of 3D medical image registration are detaileddiscussed, which include the 3D spatial transformation, registration measurecalculation based on joint histogram, genetic optimization of registration parameter.A 3D medical image registration experiment system based on MATLAB is designedand implemented. The comparing research of different registration measure,interpolation methods, optimization methods, estimation methods of joint probabilitydistribution can be conducted with this system. The system can be further extendedaccording user requirement.