Image Registration Technology Research Based On Feature Points

Image registration technology is the important study task, It is to find the same space correspondence between the subject image and object image, the basis of image fusion, image rebuilding, object motion estimation and image analysis etc.In the medical image fusion field, Different imaging equipment complement one another in the shape information and function information of the same anatomy structure. For example, CT and MR can get the structure image information, Then PET acquires function image information, namely, Though CT and MR images can display anatomy structure, and they cannot show the function information. By contrast, PET image can illustrate the metabolism function information, not the clear shape structure information because of the low resolution image. Clinician needs the integration of different image information urgently, In order to show the structure information and function information on the same medical image. Firstly, we use image registration technology to make the same space correspondence of different modality, and then integrate this different modality image information; Finally, It can implement all kinds of image devices to have complementary advantages. So doctors get much more abundant and accurate diagnosis and treatment information. Currently, The medical image registration of anatomy image and function image is main between the CT or MR image and PET or SPECT image; In the radiation therapy, We take the CT and MR image registration and fusion technology to radiation treatment plan and evaluation, It use CT image to calculate radiation dose accurately and MR image to describe the location of tumor; In the surgery navigation assisted by computer, Surgeon locate the focus of infection on the basis of CT/MR/DSA registrated image and design careful surgical plan, During the surgery, Surgeon can track the surgery according to the virtual surgical patients, actual patients and the surgical instrument; In evaluation of treatment effect, Image registration is to compare between organ or tissue images obtained in different time in the same patient, So we can monitor the disease development and treatment process; In addition, sometimes we compare the testee's image with normal image in the same position to determine whether testee is normal, if exception, We compare the testee's image with again with some typical disease image to determine whether the testee is the same disease, and sometimes we need compare patient's image with atlas.According to the properties of geometric Transformation, Medical image registration methods consist of rigid registration and elastic registration. For the 2D image, rigid registration is to find three degrees of freedom (one rotation, two translations) for mapping the correspondence points between subject image and object image. Nowadays, the rigid registration technology has been matured in different modality image, it can achieve good registration precision and robustness and has been used to clinic. However, it is only good for the rigid body image registration. For example, the brain skull almost does not deform, so the same patient brain image can use rigid registration. Then different patients to atlas and between different patients, we should take the elastic registration. Rigid registration have limited applications in medical image registration, most clinic application need the elastic registration to describe the correspondences between the medical images. At present, we use image registration method based on spline, the elastic registration model, viscous fluid registration model and optical flow model to solve this kind of problem, By contrast, Elastic registration methods are not mature enough, various methods have certain limitation, we need to choose appropriate registration method according to the different research object, It needs further analysis and research on how to improve the elastic registration precision and speed and establishing the reasonable registration deformable model.Image registration method can be divided into two kinds based on benchmark of the characteristics of registration, one kind is based on image features registration method, The other is based on pixel similarity registration method, Registration method based on image features determines registration parameters according to the transformation relationship between important features, firstly, We extract features and search the corresponding relations between features then determine registration parameters according to the similarity measure function. It relies on image segmentation heavily, if the initial segmentation is good, Registration can get greatly simplified according to image corresponding relation. Currently Registration method based on image feature is frequently used, Firstly, It defines the corresponding feature points of two images, then keep those feature points well-distributed, finally determine the equation solution according to corresponding point pair. If we have enough points, we can often use least-square method to solve the parameters of target cost function, in order to improve the accuracy of parameters. Among the registration method based on feature point, the feature point could be the corner, line intersection, point of local maximal curvature, center point of the window with local maximal curvature, the center of gravity of closed area etc.This paper is a medical image registration method based on feature points, in the feature extraction process, the commonly used methods are Harris corner detection and SIFT (Scale Invariance Feature Transform) feature point selection. Corner detection methods are mainly based on the edge of the image contour method and based on image gray level,,method based on the image contour is largely dependent on the image segmentation and edge extraction, once the local of target to be detected changes, then May lead image segmentation and edge extraction operation failed, its scope of application has been greatly restricted, and the method based on image gray avoids these shortcomings, such as Harris corner detection,its consideration is the neighborhood pixel gray change, by calculating the point of curvature and gradient to detect the corner, not the entire goal edges. SIFT method used to extract feature points which can be used for reliable matching in an object or scene under different perspectives, those feature points for image scaling and rotation remains the same, they are robust to the light changes, noise, affine change, and keep high uniqueness. the SIFT feature vector, that is gradient orientation histogram, can well describe the characteristics of point of neighborhood gradient information, so they can keep unchanged if scale change, rotation and other factors, geometric distortion, and then the length of feature vector normalization, they can remain the same if the light changes, but in the not enough rich texture images (medical images) the SIFT method can only extract a few feature points. In this paper, Harris corner detection extracted enough feature points, and described those feature points by SIFT feature vector, and then complete the image registration.Image registration method based on pixel similarity is based on statistical correlation of the two images'gray distribution, This method is flexible, It uses image gray information directly in registration process without image segmentation, But its calculation is large, Registration accuracy and robustness is bad for multimodality images and application limitation in clinical. Currently this method is mainly applied in global affine transformation of head image and image registration automatically. Commonly image registration methods based on pixel similarity cross correlation method, Joint entropy method and mutual information method, namely taking maximization of Cross correlation and mutual information and minimization of Joint entropy as Similarity measurement respectively. Due to the gray value cannot correctly reflect the image space structure information, registration results are vulnerable to local minimum value, so the robustness is not strong. The HAMMER algorithm (Hierarchical Attribute Matching Mechanism for Elastic Registration) proposed by Shen Dinggang had achieved very good results in elastic registration of brain image, it defined a attribute vector (boundary type, image grey value and geometric moment invariants) to each pixel, Due to the attribute vector reflect the information of the neighborhood spatial anatomic structure and pixels have high degree of distinction, So registration accuracy and robustness are obviously better than the algorithm based on image gray value information. Insufficient, however, HAMMER algorithm relies on the segmentation of brain results and the computation is large.The result of image registration will influence subsequent image processing work, On the basis of understanding the image registration method, this paper mainly studies the elastic registration based on image feature points and proposes a new feature description vector, and this feature description vector is composed of image grey value, type of feature points and gradient orientation histogram. Feature points have three types of corner point, edge point and flat point, Corner points are local important features, They are the pixels of the most dramatic brightness change and the maximum curvature of two dimensional image curve, with the advantage of rotation invariant and not changing if light conditions change, The corner points are detected by Harris corner detection algorithm in this paper. Gradient orientation histogram blended gradient information of all neighborhood pixels; it can describe local structure characteristics of image space well. Meanwhile, we build a new cost function and take hierarchical optimization strategy to solve the minimum of cost function; the key idea depends on registration processing with number of feature points dynamic increasing, In order to ensure the robustness of the algorithms and registration precision. The method proposed by us have tested on MR brain images and standard natural images, The experiment result shows that this method has better registration precision than the traditional methods significantly.