Muti-modal Medical Image Registration Based On Feature Matching

Medical image registration is defined as that at different time, different angles and multi-modal medical images acquired by different devices, after a series of transformation, the same anatomical structure (or points) aligned at the same space coordinate position. Image registration is a core technology of medical image processing research. It can be used for correction of medical images, and detecting changes in the anatomy for the image fusion of different modalities. Thus, in medical diagnosis treatment is widely used. In recent decades, the technology research has been more and more attention from scholars.In this paper, based on feature points matching multi-modal medical image registration has been the focus of research to make greater intensity inhomogeneity and noise medical images in a complex environment can be obtained by automatic registration. This paper includes the following aspects:1. This paper generally described the feature matching based on multi-modal medical image registration step, introduced algorithms of feature detection, construction descriptors, feature matching, parameters estimation, resampling and interpolation, however we focuses on the common feature extraction algorithms; At last, we analysed the Laplace of Gaussian extracting feature points invariance principle in terms of illumination, rotation, scale and so on.2. Between the pixel values of different multi-modal medical images for very different, leading to registration of high computational complexity, low registration accuracy. This paper presents a fusion entropy graph theory based on feature point matching registration algorithm. The algorithm uses Laplace Of Gaussian feature points that extracted by local texture feature construction weighted CS-LBP descriptor feature vectors and calculate the Renyi mutual information that estimated by generalized nearest neighbor Euclidean distance, then treat it as an objective function. Finally, using Powell optimization algorithm to optimize the objective function, and seek transformation model parameters. This method does not require the registration processing for each pixel of the image to extract more uniform distribution, stable and easy to extract features of the image for describing, between the registration process as long as these features can be calculated, the method not only reduce effects of noise and difference caused by image intensity, but also reduce the amount of calculation and registration time. The results showed that:under the premise to ensure that the registration rate, the registration precision can reach sub-pixel, and the noise can be a significant medical image registration to achieve good results.3. Based feature extraction for medical image registration algorithms, it can not take full advantage of the features of the neighborhood built by descriptors, this paper presents a new texture-based feature descriptor construct by Local Wavelet Pattern(LWP) and combined probability statistics based on consistency point drift algorithm (Coherent point drift, CPD) named LWP-CPD algorithm, which makes full use of the spatial relationship with the neighborhood and the center of the neighborhood and the neighborhood each other, then use the maximum likelihood estimation motion consistency and velocity field panning feature points, and ultimately aligned position in space. By comparing the experimental results, the algorithm registration speed, accuracy and robustness are much better than the SIFT-CPD and traditional NMI registration algorithms.