Research On Image Representation Based Non-rigid Multi-modal Medical Image Registration

With the development of imaging technology,the emergence of various imaging devices has made great contributions to the advancement of modern medicine.However,due to the limitation of imaging principles,single-modal imaging technology can only provide single and limited information.Therefore,to improve the accuracy of diagnosis and the effectiveness of treatment,doctors often need to fuse the information from different modal images to obtain the comprehensive information of diseased tissues or organs.The key of multi-modal image fusion is accurate image registration,but it faces many challenges.On the one hand,in clinical applications,most human organs are soft tissue organs.The patient's breathing,body position changes and the force of imaging sensors(such as ultrasound probes)may lead to complex non-rigid deformation of soft tissues,which will result in more parameters and greater computational burden in registration processing.On the other hand,due to the difference in the principles of different modal medical imaging,there is no simple one-to-one correspondence for the intensities of the different modal images,which makes it very difficult to accurately construct similarity metrics for non-rigid multi-modal medical image registration,and it will affect the accuracy of the image registration eventually.The-state-of-the-art non-rigid multi-modal medical image registration methods include the intensity based and geometrical feature based method.However the geometric feature based methods largely rely on the accurate extraction of geometric features such as points,lines and surfaces in the image which is still an open and challenging problem.The intensity based registration method mainly includes mutual information based method and image representation based method.The former is of high degree of automation,and does not depend on the correspondence of image intensities.However,it does not take spatial structure information in the image into account.In addition,it is easy to fall into the local extreme value,which will result in misregistration.The latter is to transform the multi-modal medical image registration problem into a mono-modal registration problem through a certain mapping relationship,and then utilizes the similarity metric in themono-modal image registration to realize image registration.Compared with the mutual information based method,the image representation based registration method can construct the similarity measure(such as SSD measure)more easily,and the considered information is more comprehensive(including intensity and structural features).In addition,this method finally solves the mono-modal image registration problem,so it has the advantages of simple calculation,and its optimization is not easy to fall into local extremum.These advantages render this method to be more potential to solve the non-rigid multi-modal medical image registration problem than the mutual information based method.Typical examples of image representation based registration methods include the Laplacian Eigenmaps method,the entropy SSD method,and the modality independent neighborhood descriptor method.The Laplacian Eigenmaps method is sensitive to noise,the entropy SSD method has a fuzzy structural representation result and the modality independent neighborhood descriptor(MIND)is not of rotational invariance.Meanwhile,these methods use the artificially designed feature extraction methods to achieve image structure representation,so they are difficult to accurately describe information for the complex medical images.To address the above shortcomings,this dissertation will start from the view of image representation to do in-depth research on the more accurate and more robust non-rigid multi-modal medical image registration method.The main work of this dissertation includes:1)In view of the shortcomings of the entropy ESSD method such as the fuzzy representation result,the existence of pseudo-points and the sensitivity to the size of image patch,a spiking cortical model(SCM)and fractional order generalized entropy(FOGE)based non-rigid multi-modal medical image registration method is proposed in this chapter.This method uses the multiple ignition information of SCM to dig the image feature information instead of simply calculating the intensity distribution information of the image patch,and then uses the more sensitive FOGE to construct a new entropy inage to represent the different modal images.This method can provide a more clear and consistent entropy image for image registration.Experiments show that this method can achieve better registration results than the traditional Entropy SSD(ESSD)method and the classical normalized mutual information(NMI)method.2)To address the problems that the traditional entropy SSD method is sensitive to noise,the self-similarity calculation of MIND method is not accurate enough and MIND is not rotation invariance,this chapter proposes a Zernike moments based local descriptor(ZMLD)which is inspired by self-similarity for image registration.The method first calculates the Zernike moment of image,and then constructs the ZMLD based on the self-similarity of the moment feature.Based on the SSD of ZMLD,the similarity metric is computed for multi-modal image registration.Finally,the non-rigid multi-modal medical image registration will be realized based on the above metric.Compared with the MIND method whose computation of the self-similarity only depends on the image intensities,the proposed method in this chapter takes the intensity information and spatial structure information into account,which makes the calculation of self-similarity more accurate and the ZMLD rotationally invariant.Experiments show that this method has provided greater accuracy and better robustness compared with the other state-of-the-art image representation based registration methods.3)To overcome the limitation of the artificially designed feature extraction technology in the existing image representation based registration method,this chapter proposes an unsupervised Laplacian Eigenmaps(LE)based deep learning network(LENet)for image registration.The LENet is first used to obtain the various levels of image feature information.Then the self-similarity of these feature information is utilized to construct the learning based data-adaptive descriptor(LDAD).Finally,the SSD of the LDAD is used to perform the accurate non-rigid multi-modal medical image registration.The off-line learning procedure of this method is highly simple.In addition,the LDAD can adaptively extract the essential features of the image as the data changes,which ensures the efficiency of feature extraction and image registration.Experiments show that the LDAD method can provide better robustness and higher registration accuracy than the NMI method,the ESSD method,the Weber descriptor method,the MIND method and the ZMLD method.