Non-rigid Registration For Medical Images

Medical image registration is a key component in medical image processing andanalysis, including image comparison, image fusion and image variation analysis andtarget recognition. Medical image registration is now applied in image fusion, surgicalnavigation, response evaluation, pathological tracking, auxiliary medical diagnosisand radiotherapy planning. In this dissertation, we concentrate on the non-rigidregistration algorithm based on image intensity, feature points and mixture algorithmto improve the veracity, accuracy and topological homeomorphism. GPU is used toaccelerate the non-rigid registration. The major contributions of this dissertation are asfollows.A diffeomorphic Demons algorithm with topological homeomorphism based onRiemannian manifold is proposed due to the traditional diffeomorphic Demonsalgorithm lacking of topological homeomorphism. A three dimensional medicalimages is treated as a four dimensional manifold in four dimensional Riemannianspace by Sochen-Kimmel-Malladi Nonlinear diffusion equation. So, the three imageregistration is transformed to a four dimensional curve evolution. In order to correctthe topology change caused by intensity offset, the deformation field and the intensityoffset are computed at the same time by minimizing the Polyakov functional. Asegmentation result is used as the prior knowledge to constraint the deformation fieldenergy and strengthens the topological homeomorphismA robust non-rigid registration approach based on the Students’t-distributionmixture model is proposed due to the Gaussian mixture model being vulnerable by theoutliers and the data with longer than normal tails. The Gaussian mixture model is a special case of the Students’t-distribution mixture model in theory. The parameter setof the Students’t-distribution mixture model is solved by EM algorithm. The weight ofeach float point is calculated in EM algorithm instead of a constant in order to reducethe impact of outliers. The degree of freedom of each point in theStudents’t-distribution mixture model is calculated to change the probability densitydistribution to avoid estimating the noise level of data sets in the Gaussian mixturemodel that may bring the additional error. We impose the local spatial constraints byadding the Dirichlet distribution and enhance the accuracy and anti-jammingcapability of the Students’t-distribution mixture models. The points have a feature ofcoherent point drift by adding regularization into the expectation function. Moreover,we extend the algorithm to rigid registration and affine registration.In order to solve the problem of the aliasing transformation field in medicalimages registration, a hybrid non-rigid registration algorithm including feature pointsand intensity is proposed in this dissertation. A double-layer correction approach isused to revise the aliasing transformation field produced by the diffeomorphicDemons algorithm. Firstly, the displacement vectors produced by feature pointregistration are used to pre-correct the field. Secondly, the diffeomorphic Demonsalgorithm with an improved regularization is used to align the corrected image. Theupdate step length is selected adaptively according to the distances between the pixelsand the interval region. The improved diffeomorphic Demons algorithm not onlyimproves the accurate in the target regions but also improves the veracity in theovercorrection regions.The non-rigid registration is slow due to a large number of control points, theiterative strategy and the normalized mutual information (NMI). A parallel algorithmcombining a B-spline coefficient optimization algorithm is proposed to accelerateregistration. In this algorithm, the data parallel algorithm computes NMI and the taskparallel algorithm, which the data parallel algorithm is embedded in, computes thegradient descent flow. Control points are restrained on the targets by computing imagelocal entropy to reduce computation. A thread balance algorithm based on Greedy isused to solve the computation imbalance problem caused by the uneven distribution of control points.