Research Of Medical Image Registration Method Based On DTCWT And NPSO

Medical image registration refers to the space agreement of two medical images after one image's one or a series of geometry transformation. It is a new interdisciplinary research field based on information science, computer image processing and medical science, which has potential clinic benefits and a bright future in application.Dual-tree complex wavelet transformation not only maintains the traditional characteristics of multi-resolution and local analysis ability in time domain but also possesses other properties, such as smoothly shifting invariance, good directional selectivity, limited data redundancy and perfect reconstruction. Its time-frequency feature is also superior to that of the traditional discrete wavelet transform.Based on the thought of birds of a feather flock together, niche particle swarm optimization algorithm is mainly used to solve multi-peak functions optimization problems. It can efficiently improves the situation of easily being trapped in local optimum, low searching efficiency and poor optimum performance during particle swarm optimization algorithm's searching process.After a systematic study on the algorithms of wavelet, dual-tree complex wavelet particle swarm and niche particle swarm optimization, this thesis proposes a novel medical image registration algorithm which is based on the combination of dual-tree complex wavelet and niche particle swarm optimization. The threshold hausdorff is taken as similarity measure to make two sets of points matched. Niche particle swarm optimization algorithm is used to search for the optimal transformation parameters. Firstly, pre-processing is realized for two images, including noise reduction and extracting the multi-scale key point by dual-tree complex wavelet transformation. Secondly, an overall coarse registration is implemented by affine transformation. Thirdly, a local fine one is finished by B-spline elasticity model.This thesis is based a lot of experiments, including registering respectively the images of SPETCT and MR, CT and MR, MR_T2 and MR_PD. Simultaneously, registration results are compared in aspect of feature extraction, optimal algorithm and transformation model. Evaluating parameters such as correlation coefficient, minimum mean square error, normal mutual information and signal-noise ratio are used to assess objectively the registration results. Judging from the evaluate results, the conclusion can be drawn that the proposed medical image registration is superior to the traditional ones in the aspect of accuracy and robustness.