Research And Application On Fast Registration Methods For Abdominal Medical Image

Medical image registration is a research topic about the information processing, computer image technology and modern medicine and other disciplines, which has been extensively used in clinical diagnosis, treatment and preoperative planning. Medical image registration refers to match and superpose two or more time-varying images (such as portal venous phase and venous phase of abdominal CT images, etc.) or heterologous images (such as MRI, CT, PET, SPECT), using computer digital image processing technology. Usually, image registration is treated with single or multiple space transform, making its feature points and the corresponding feature points of reference image to achieve spatial coherence. In the clinical application, image registration is one of the basic medical image processing methods which is the foundation of the subsequent operations such as image analysis, fusion, or 3D visualization.With the development of diagnosis and medical information technology, the real-time requirement of medical auxiliary method for clinical medicine to medical image registration becomes very important. Therefore, how to achieve real-time medical image registration in condition of the accuracy of registration has become a hot issue in the research field. However, with the development of medical imaging technology, the accuracy of medical image is more and more big. We cannot simply rely on high performance computers to solve performance problems in image registration, and should be studied from the aspects of image registration algorithm, so the traditional method is facing new challenges. With the characteristics of organs softness, respiratory motion and strong deformation, it is impornt to research abdominal medical image registration.Medical image registration methods mainly include 3 categories:they are the image registration algorithm based on gray information, the image registration algorithm using features information, and the image registration algorithm using biomechanics. The main innovations of this article are introduced around these 3 categories:Firstly, a MRAI based fast 3D medical image registration methods proposed. Firstly, it adopts the free deformation transformation of a two stage registration algorithm to obtain the preliminary result. To obtain more accurate results, alternating iterative registration algorithm is employed. To reduce the alternate iterations of the registration algorithm, amulti-resolution analysis method is used. In order to further accelerate the algorithm, it adopts a high performance computing method based on CUDA(Compute Unified Device Architecture). It make full use of the advantage of parallel computing of GPU(Graphic Processing Unit)in the CUDA framework, which is used for the abdomen three-dimensional CT images quickly registration by combing with image multi-scale, maximum mutual information methods.Secondly, a Scale Invariant Feature Transform(SIFT) fast calculation method based on the unified computing equipment structure is proposed to solve the efficient problem of the extraction of SIFT features in calculation process. The new algorithm makes full use of the advantage of image processing unit in parallel computing, floating point calculations, memory management, and other areas. The speed of the extraction of SIFT features is accelerated by CUDA framework, and the proposed scheme balances the resource distribution of CPU and GPU. Furthermore, the new algorithm re-designs the process of the implementation of the algorithm, which lead to a faster calculation speed.Finally, a fast liver CT images registration method based on the biological mechanics model is put forward, which is accelerated by CUDA. Firstly, the grid processing is applied to the original image, to obtain the significant edge information of image by adopting three normalized reverse distance transform. After that, a new energy function is employed to constrain registration process, and a minimize energy function is used to perform image registration. The energy function contains a kind of internal strain potential energy and two external energies based on gray level and characteristics respectively. During the calculation of the numerical solution of the energy function, the finite element is used to accelerate the level set evolution process, the CUDA framework governs the process, which lead to the faster liver CT images registration.All the data in the experiments are from actual clinical abdominal CT images and are verified in the clinical trials:MRAI abdominal CT three-dimensional medical image registration based on CUDA accelerated can guarantee registration accuracy, which can be quickly performed. Compare to the traditional 3D image registration algorithm based on CPU framework, the proposed scheme is 255 times faster than the original one; Abdomen medical CT image registration based on CUDA accelerated SIFT feature extraction greatly improve the SIFT feature extraction operation speed. The acceleration ratio increases along with SIFT feature points increased and maximum acceleration ratiocan reach 19.54 times. By Appling the SIFT in medical image registration, the algorithm can perform accurate registration quickly; the biomechanical liver image registration method based on CUDA can make abdominal CT image registration quick and accurate. The speed of the proposedalgorithm is 46.8 times faster. In the process of building CUDA thread,192 threads in every thread block will get higher efficiency and better performance.The proposed fast abdomen medical image registration algorithm is able to provide a rapid and real-time multi-periods abdominal CT image registration for clinical computer aided diagnosis. It can realize liver and other organizations registration of two CT images from different periods, so as to improve the doctor’s diagnosis process and improve the diagnosis and treatment level.