Improved Sift Algorithm And Its Application In The Medical Ct Image Retrieval And Registration

Medical CT image retrieval and registration are the current research focuses of medical image processing and analysis. They are also meaningful for clinical diagnosis and treatment. Scale Invariant Feature Transform (SIFT) as a very effective feature detection and matching algorithm, is invariant to the rotation, illumination, scale and affine transformation. With these advantages, it can be used in content based medical image retrieval and feature points based medical image registration pretty well.The main work of this dissertation is to improve the SIFT algorithm, and make a thorough research of its application in medical image retrieval and registration. Specifically, it includes the following parts.Firstly, considering the situation that SIFT algorithm may cause a number of wrong points matching in the actual application, we propose an improved algorithm based on two sizes of overlapped blocks and gradient decomposition. By using two kinds of size, partly overlapped small regions to calculate the gradient direction histogram, and decomposing the gradient, the feature descriptor generation method is improved. The experimental results show that for cases with the noise, rotation, scaling and so on, this improved algorithm has better robustness, more matched point-pairs, and higher matching accuracy.Secondly, aiming at the problem that large amount of calculation and long retrieval time when using the SIFT features in medical image retrieval, we propose two methods by using approximate nearest neighbor (ANN) algorithm and bag of words (BoW) model. One is CT image retrieval using SIFT features and approximate nearest neighbor algorithm, and the other is CT image retrieval using SIFT bag of words model. The experimental results show that these two methods greatly improve the retrieval speed, and nearly have no influence to the retrieval accuracy. Compared with the traditional texture-based retrieval method, these two methods have higher precision and more capability to retrieval similar images with the detail features of the target image. Therefore, they may have the application value in large-scale medical CT image database retrieval, and research potential in other types of medical image retrieval.Thirdly, considering that it may cause insufficient precision and adaptability only using the matched SIFT feature point pairs to estimate the transformation parameters, we propose a medical image registration method based on SIFT and edge points matching. Here, the feature points of the reference image and the float image are extracted and matched by using SIFT algorithm, and the preliminary registration result is obtained by estimating the transformation parameters. Then, the edge points of the reference image and the preliminary registration result image are extracted respectively, and the Coherent Point Drift (CPD) algorithm is used to register these two points set. Finally the accurate registration result is obtained. The experimental results show that this method can achieve accurate registration results for medical images with the complicated deformation, and is expected to become the basis of medical image fusion technology.