Research On Technology And Application Of Medical Image Registration

Image registration plays an important role in medical image analysis.With the help of this technology,doctors are able to combine image data of the objects taken at different time or by different sensors to obtain more complete information about the patient.In general,the common image registration approaches can be roughly divided into two categories: the feature-based methods and the intensity-based methods.Specific regions,lines,points and so on in the image all can be treat as features.As extracted features are only part of the image,the matching process based on features is relatively time-saving.However,small number of features also brings disadvantages.For example,even a small proportion of extracted features are outliers,it still will have large influence on registration results.So,the matching consequences derived from feature-based methods are usually low-accurate and low-precision.On the contrary,the intensity-based methods show more accurate than feature-based ones because they take the whole image into consideration,but the calculation of all pixels in the image is really time-consuming.In addition,the intensity-based methods used to consider the pixels of the image as isolate individuals while ignoring the spatial correlation between them,so these approaches have low robustness to disturbance.Aim at partially or totally solving the shortcomings of these two kinds of methods mentioned above,this paper conducted the thorough research of related problems and obtained some achievements as follows.First,this paper proposes a new hybrid registration measure,called contour and neighbor volume similarity(CNVS)method,which incorporates merits of both area-based and feature-based methods.The implementation of this integrated method can be illustrated with a coarse-to-fine registration framework.In the coarse registration stage,the closed contours of the objects are first extracted as a stable feature set.Based on a distance measure,the feature set is used to rapidly estimate an initial transformation between reference and floating images in the global scope.Subsequently,in response to the possible false alignment when registering symmetrical objects with feature-based methods,we employ an alignment correction procedure to ensure the reliability of the original transformation.Finally,the modified feature neighborhood and mutual information,an area-based method characterized by multiscale filtering mechanism,is adopted in the fine registration stage to obtain a precise final transformation.In addition,we introduce an equilibrium strategy to the differential evolution(DE)algorithm for estimating transformation parameters in the coarse registration stage.Our proposed method has been extensively evaluated by comparing with several state-of-the-art registration approaches on multi-modal brain images.The results indicate that it can automatically align images in different noisy environments with high accuracy and robustness.Then,the paper presents a spatial structure descriptor for image representation,which facilitates the registration of multi-modal image pairs.The construction of this descriptor is only based on structural and spatial information of the image without involving the intensity correlations,so it is modality independent.In particular,to exploit structural information,we calculate optimal combination of low and high order differential derivatives of the image in the gauge coordinate system.Experimental results have shown that the sum of absolute difference(SAD)of the descriptor as a similarity method performs more robustly and efficiently than other well-known similarity measures.