Research On Group Matching Method Of Multispectral Fundus Images

Vision is an important sense of human,and fundus disease is one of the most common visual disorders.Fundus retinal images can be used for examination of various fundus diseases and early diagnosis of systemic diseases.Fundus image registration is a key technology in retinal medical image analysis.Multispectral fundus imaging is a recent invention of non-contact,high resolution fundus coronal multimodality imaging technology,using several different wavelengths of spectral band through fundus retina and choroid to get the different depth of the fundus sequential slices.The non-invasive retinal and choroidal images provided by this technology can more highlight the anatomical structure and metabolic information by integrating spectral information.The emergence of multispectral technology is an innovation of fundus disease diagnosis technology.However,the acquisition time of multi-spectral fundus images is often longer than the natural eye scanning,which often leads to the spatial deviation of the corresponding physiological structure in the multispectral fundus images,and the difference in the anatomical structure between different sections makes the analysis more difficult.To effectively solve the spatial misalignment and quantitatively analysis the anatomical information between multispectral slides is conducive to assist doctors to make accurately medical diagnosis and monitor the development of the disease.The realization of multispectral image registration has important clinical practical value for ophthalmic diagnosis and treatment.At present,the fundus image registration methods are usually based on the pairwise registration.Due to the potential correlation between multispectral image sequences,the use of correlation information between sequences is ignored in the registration based on any two images.The existing groupwise registration method is mainly applied to the single-modality images registration,and most of these methods rely on the pairwise registration results.Aiming at the above problems,this paper proposes multispectral fundus image groupwise registration methods,designs groupwise registration framework based on semidefinite programming and quadratic programming that comprehensively considers the image sequential information,and obtains the feature point matching relationship between all sequential slices via joint optimization.Experimental results on the fundus image data captured from patients and healthy subjects show that the proposed groupwise registration method has a significantly higher accuracy than the pairwise registration.Moreover,this method is robust to noises,and even when the corresponding relations of feature points are missing due to different anatomical structures between sections,it can still show better matching results.The main innovations and contributions of this paper are as follows:(1)The early fundus image registration algorithms are only applied to the pairwise registration.In this paper,a novel multimodality image registration method based on groupwise registration is proposed,which combines multispectral image sequences to match feature points simultaneously,so as to make the registration results closer to the optimum solution and improve the accuracy of registration.(2)The corresponding relation of feature points designated by medical experts can be incorporated into the proposed registration method framework in the form of constraints,allowing medical experts to manually intervene in the registration process with their rich experience,and the proposed framework can be executed automatically or semi-automatically,which is beneficial to the clinical application of ophthalmologists.(3)Due to the noise in the process of multispectral fundus image acquisition,as well as the inevitable errors in the calculation of feature point matching cost,the proposed method can correct the mismatching relationship caused by noise through the correspondence relationship between the identified pairs of feature points,and it has high robustness when the quality of fundus image of patients with fundus lesions is poor,resulting in the loss of feature points.