Research Of The Non-rigid Registration Method For The CT Mice Images

As the development of the new imaging technology, it is very important for vivosmall animal in the preclinical study. Small animal is also the important experimentalobject for the modern biomedical research. The study of small animal image registrationis essential to the multimodality image fusion, to segment the organs, to assess responseto therapy. In this dissertation，the non-rigid registration methods for the mice imageswere studied.It is complex for the mouse’s body with it composed of bones and soft tissues. Thenon-rigid deformation of bones and soft tissues may be different in different postureswhich will influence the accuracy of mice image registration. One of the useful methodsfor mice images registration is sub-step. But most of those methods will consume lots oftime as the size of the mice image is large. It is meaningful to reduce the registrationtime and guarantee registration accuracy at the same time.Based on the existing registration methods, this paper presented a new registrationmethod for mice images. It included three steps: the first step is preprocessing of themice images. The mice would be rotated and transformed for correcting themismatching, which were caused by the different initial scanning position; the step twois the registration based on the feature points. In this part, for each axial slice in theskeleton image, the center of each of connected areas was located and the set of centralpoints was used to be the feature points. Then the skeleton of the mouse was broughtinto approximate correspondence with a point matching method called TPS-RPM (therobust point matching algorithm based on thin plate spine). The thin plate spinetransformation computed based on the points was applied to the entire mouse image.The step three is the precise intensity-based non-rigid registration method. Multi-resolution and random sampling strategy was used to increase the speed of registration.The registration progress would use normalized mutual information as a similaritymeasure, B-spline transformation as a transformation model. To reduce the registrationtime, pseudo-random sampling algorithm would be used to computing the similarity ofthe image. Adaptive stochastic gradient descent algorithm would be used to optimize thetransform parameters. The program of this step was based on elastix which is a usefultoolbox for medical image registration.This paper evaluates the proposed registration method with Micro-CT datasets of mice. These experimental results demonstrate the effectiveness and robustness of thismethod.