Registration Of The Same Modality And Multi-modality Fundus Images

There are many kinds of fundus diseases,such as retinal vascular diseases,inflammatory diseases,macular lesions,retinopathy,and so on.Fundus diseases affect people's daily life,and cataract,glaucoma,age-related macular degeneration,diabetic retinopathy is the four major cause of blindness,to people's daily life has brought great inconvenience.Fundus image is the main method used to diagnose the disease of fundus oculi at present,red free fundus images(RF)and fundus fluorescein angiography(FFA)are the two most widely used images.The visual field of an image is limited,the registration of the same modality fundus images can be used to synthesize a larger field of vision to determine the location of the lesions.In addition,the information of the red free fundus images and the fundus fluorescein angiography is complementary,the registration of multi-modal fundus images can provide more accurate information for diagnosis.However,at present,the registration of these two kinds of images is realized by the "artificial silhouette" of the doctor's mind,which requires a high level of the doctor,and the diagnosis time is long,the probability of missed diagnosis and misdiagnosis is relatively large.In order to study the registration of fundus images,this paper studies the theory of the image registration,analyzes the advantages and disadvantages of the existing image registration algorithm.The registration of fundus images can be divided into the same mode of fundus image registration and multi-modal fundus image registration.In this paper,the classic image registration algorithm SIFT is used for the registration of the same mode fundus images,and the relevant parameters of the SIFT algorithm are adjusted,so that the SIFT algorithm can be applied to the lower contrast fundus images.However,the gray level difference between the multi modality fundus images is nonlinear,and the SIFT algorithm is no longer applicable.In this paper,according to each step of SIFT,we analyze the specific reasons that SIFT algorithm is not suitable for multi-modal fundus image registration.For the reason of the analysis,this paper proposes a new method for multi-modal image registration based on rotation invariant distance.First extract relatively stable feature points of fundus vessel with Harries,then build a rotation invariant descriptor for each feature point,then calculate the rotation invariant distance between feature points,matching feature points according to the rotation invariant distance,the last,RANSAC eliminates the false matching,estimates the transformation matrix and realizes the image fusion.This paper studies the fundus image data are from the Nanshan hospital,including 80 pathological red free light fundus images,80 pathological fundus fluorescein angiography,80 healthy red free fundus images and 80 fundus fluorescein angiography,a total of 320 pieces of fundus images.The mean square error(RMSE)is used to analyze the registration error of the same mode of fundus image registration and multi-modality fundus image registration.According to the analysis of experimental data,the error of fundus image registration based on SIFT is 2.40±0.5,the error of proposed multi-modal fundus image registration based on rotation invariant distance is 0.89± 0.72,matching the success rate is above 92.4%,and the performance of the algorithm is not affected by the image rotation.It is proved that the classical SIFT algorithm can be applied to the registration of the same modality fundus images,it is also proved that multi-modal image registration based on the rotation invariant distance proposed in this paper overcome the nonlinear gray difference of multi-modal fundus images,realize the multi-modal fundus image registration.