Research On Some Key Techniques In Medical Image Registration Based On Features

Medical image registration, a crossing research topic of information science, computer image technology and modern medicine, has been applied widely in the areas of clinical diagnoses, therapy and preoperative planning. This dissertation comprehensively analyzes and summaries the related concepts, methods and techniques of medical image registration, and the feature-based medical image registration is emphasized. The feature is the easily-recognized, geometric and anatomic information element in the image, for example, point, line, contour etc. Under the guide of the features, image registration can be implemented intentionally and quickly. For the practical problems in medical image registration, some achievements as following are obtained.Aiming at the condition that in medical image, the regions that are different or cared are restricted to specific domains, the application of radial basis function with compact support to medical image registration is studied. A local image registration method, the combination of feature points and intensity information to optimize the transformation function, is proposed. Each feature point has an influence radius that can control the action domain of the feature point. For the difference of the intensity between monomodality and multimodality image, the sum of squared difference and mutual information are respectively chosen as similarity measures that can evaluate the result of image registration. The transformation function is initialized by interpolating the feature point-pairs, and then optimized it with niche genetic algorithm (NGA). In comparison with GA, NGA effectively overcomes the drawbacks of premature and weak exploitation capabilities and the precise image registration results are achieved.Due to the limitation of image registration with one kind of feature, multiple features should be introduced to guide image registration. This dissertation proposes an elastic medical image registration method using both contour and feature points. The linearization is employed to reduce key points in the extracted contour, which improves the computation efficiency. The sum of the distances between feature point-pairs and the distances between contour-pair is regarded as the criterion of image registration. The transformation function is not resolved iteratively until the alignment of feature point-pairs and contour-pair is achieved. The presented algorithm simultaneously registers the global structure information of the image (contour) and the local detail information (feature points), so that the difference between images can be compensated exactly.A medical image registration algorithm using feature points and curves is proposed on the basis of the research about the application of the feature curves to imager registration. This presented algorithm not only has the accuracy of image registration based on points but also possesses the robust of image registration based on curves. The feature point-pairs and feature curve-pairs from the images can be exactly extracted under the help of semi-automatic extraction method. The extracted curve-pairs are modeled by non-uniform cubic B-splines so that they have the same parameter space. The mechanism of non-uniform subdivision of feature curves ensures that the discrete points can match the original feature curve as close as possible and fulfill the requirement of image registration. The optimal transformation function, which is initialized by interpolating the feature point-pairs with thin plate splines, is solved iteratively by continuously improving the regions of maximum difference between two feature curves. The introduction of feature curves can implement the registration of the continuous information in the images.A new feature curve based local image registration method, which can make a smooth deformation for feature curve and its neighborhood, is proposed based on the benchmark of corresponding relationship between feature curve-pair. The feature curves extracted from the images are modeled by non-uniform cubic B-splines, and the action domain of each feature curve is divided into two parts: inner action domain and outer action domain. In inner action domain, the local coordinate of each point related to test feature curve need to be calculated, and then the point is deformed by finding the position with the same coordinate related to reference feature curve. On the basis of precise alignment of the boundary, some corresponding point-pairs are distributed as evenly as possible in outer action domain. Outer transformation function, which is built by interpolating these point-pairs, is used to register the image in outer action domain. In proposed algorithm, the deformation of the image remains continuous and local by building the same outer boundary for test feature curve and reference feature curve.