The Research On Non-rigid Registration Of The Medical Image

Image registration includes two types:The rigid registration and the non-rigid registration. At the present time, most of the research methods are based on the rigid. However, in medical field, the majority of the object of study is the non-rigid, including the deformation which is caused by the growth,the surgery,the illness or the individual difference,the distance between the points on the image is no longer unchanged. The great difference between the rigid and the non-rigid registration depends on the space conversation's essence:the rigid registration's purpose is to search the corresponding points which are on the aim image from the source image,the rigid registration's transformation is linear.Exactly opposite, the non-rigid registration's transformation is nonlinear.Two MRI brain images are chosen as the research object in this paper. The registration process this paper discusses is based on the two medical images.Firstly, the B-Spline Function is adopted as the registration method, which divides the MRI brain image into 8*8 blocks, whose principle by establishing a basic function to control the displacement of the points on the image so as to fitting every displacement of every point on this image. After the B-Spline blocks dividing,in order to facilitate the calculation, we adopts the mean square deviation as the similarity measurement.At last, the Powell algorithm is put into use as the optimal algorithm, which has a high searching speed and the local searching capability. By being displayed on the 3D Slicer platform, it can be seen that owing to the single density of the B-Spline blocks,the result of the images registration is not ideal.Secondly, another registration method is used in this paper:normalized mutual information. This method is only based on the gray level of the image. Considering the correlation of histogram, it can get the high registration precision which is the sub-pixel. The normalized mutual information is based on the ratio of the combined entropy and abundance entropy, which makes getting the optimal result easily. As the Powell algorithm relies on the starting point, which makes the process of the registration sink into local optimal easily, the Genetic Algorithm is chosen as the optimal algorithm.Experiment shows that a large portion of the image has achieved a good result. However, as every time of the calculation is based on the whole area, it has less capability of the local disposal, which makes the result of the local registration less satisfactory.Lastly, a new method is put up which combined with the B-Spline and the normalized mutual information. The image's general registration is based on the mutual information.Then the local detail registration is based on the B-spline. The experiment shows that this method can not only improve the unsatisfactory state of the B-Spline, but also can resolve the inaccuracy caused by the normalized mutual information. This algorithm is the ideal one.