Non-rigid Registration Of Lung Surfaces Based On CT Images

Lungs are one of the most important organs of the human respiratory system. Lungdiseases have become a chief threat to human life and health. With the rapiddevelopment of medical imaging technology and its combination with computertechnology, areas such as computer-assisted diagnosis and treatment, operation guidingsystem, operation teaching and training have been enjoying vigorous development andplaying increasingly important roles in the maintenance of human health. Among theexisting medical imaging techniques, Computed Tomography (CT) technology is widelyused in medical imaging of lungs. For the same patient, due to the factors of medicalequipment’ interference, breath and growth of lung, CT images obtained at differentmoments are not the same. Because compensating the deformation of lung is importantduring the implementation of computer-aided diagnosis, operation process and so on,registration of lung surfaces has become a key technology.Data of lung surfaces are needed before applying registration of lung surfaces. Inthis paper, a contour extraction method with a high degree of automation andaccurateness is proposed. At the beginning, Gauss filter is applied to the lung CT imageto remove noises; OSTU algorithm is used to select threshold for converting lung CTimages to binary images. And then, most of the unrelated areas are removed bymathematical morphology methods; Thoracic boundary can be obtained with boundarytracking algorithm and this boundary is filled by flood fill algorithm. By performingimage binarizing process and extracting thoracic boundary, we get lung contour; After3D reconstruction of lung contours extracting from lung CT image series,3D lungsurface is well prepared for image registration.This paper introduces the iterative closest point (ICP) algorithm, and puts forwardtwo kinds of methods to improve its performance: principle component analysis for datareprocessing; improving the stop condition of this algorithm. This paper analysizes thethin-plate spline function for image registration. Optimization method in registrationplays a crucial rule. This paper discusses the deterministic annealing algorithm and laida theoretical foundation for the application of this algorithm. This paper introduces thefuzzy correspondence matrix to solve the problems of noise and redundancy in featuresets. Based on the above research, this paper presents a feature based nonrigid imageregistration algorithm. The algorithm uses thin-plate spline function to parameterize nonrigid transformation, adopts the fuzzy correspondence matrix to solve the featurepoint matching problem. They are both embedded into the framework of deterministicannealing algorithm to achieve the optimal value of registration. Experimental resultsdemonstrated the effectiveness of this algorithm and showed strong robustness and highregistration accuracy compared with the improved ICP algorithm. Finally, we used thisregistration algorithm for registration of lung surfaces and achieved a good result.