Non-rigid Medical Image Registration Methods

Medical image registrations which can map points (or surface) from one medical image to homologous points on an object in the second image is one of the fundamental tasks in image processing, and then fuse them into a new image to provide doctors more rich clinical information, making all kinds of video equipment in the information complementary advantage. Due to the diversity and complexity of the medical image, which method of image registration can be chose and applied is depend on the problem. Proper registration methods have to be selected to aim at specific subjects investigated and considered to complement each other in terms of maximizing favorable factors and minimizing unfavorable ones.Based on many important clinical applications need to describe the spatial of the image using the non-rigid transformation, this thesis researches on some methods of non-rigid registration and improves them, then analyses both the precision and efficiency of registration. The concrete studies are as follows:Firstly, non-rigid registration based on pixels has been researched in the aspect of registration precision. Under the framework of differential, by using the model of local affine transformation and global smoothing constraint, it captures the nonlinear distortions, simulating the elastic transformation of the medical image. The simulation experiment shows that the algorithm is practical and the results are analyzed, highlighting the problem of low efficiency of registration.Secondly, according to time-consuming, the model of free-form deformation (FFD) based on B-spline has been studied. Based on layered and multi-resolution, normalized mutual information has been adopted as similarity measure. At the same time, image blocking which is featured with rapid calculation of local updating has been used to simulate local elastic deformation to increase the efficiency of the registration. The simulation experiments show that the algorithm is practical, and evaluation parameters, such as sum of squred difference, con-elation coefficient and time, have been used to compare with the algorithm studied above.Finally, Based on the algorithms studied above, a method of non-rigid registration which combined B-spline with pixel is formed to increase the precision of registration, besides, it inherits their advantages. So it has universal significance. According to exist the problem of low efficiency, a kind of Hermite derivative filter is introduced into the algorithm, instead of original H. Farid's optimally rotation-equivariant directional derivative kernels, and it makes full use of image information and reduces the computation. The simulation experiments show that the improved algorithm can keep the original precision in the circumstances, improve the efficiency of registration effectively.