Optical Imaging Quality Improvement Of AOSLO And High Precision Retinal Tracking Technology Based On AOSLO

Human retina is the only inner structure of the human body for noninvasive observation and high-resolution imaging in vivo.Most of retinal diseases as well as some systematic lesions will lead to the changes of physiological structures and tissues of retina.Therefore,high-resolution and high-contrast retinal images make great sense to both of early diagnosis and medical research of related diseases.In addition,highresolution and real-time imaging of the retina is helpful to continuously track and stimulate the ROI of the retina,which is of great significance for the study of human visual function and working mechanism.Adaptive optics scanning laser ophthalmoscope(AOSLO)can correct the aberrations of human eyes in real time by means of adaptive optics(AO)technique,and accomplishs high-resolution real-time retinal imaging.However,for the sake of safety of human eyes,AOSLO has strict requirements on exposure power,resulting in low contrast and low signal-to-noise ratio(SNR)of retinal images.On the other hand,influenced by human dynamic aberration characteristics,wavefront detection accuracy and deformable mirror correction ability,AO can not completely correct the aberration of human eyes,and the residual aberration will degenerate retinal images’ resolution and contrast.These factors limit the ability of AOSLO in medical diagnosis.Simultaneously,due to the complex retina vabiration,retinal images collected by AOSLO suffer from inter-frame and intra-frame distortions,which makes it impossible for people to steadily observe and track the region of interest(ROI)of retina.In this paper,a fast wavefront reconstruction algorithm based on second-order B-spline function is proposed to improve the accuracy of ocular residual wavefront restoration.The quality of AOSLO images is significantly improved by applying them with a predenoising operation and a blind-deconvolution algorithm(in which the point spread function(PSF)is estimated firstly),and it lays a solid foundation for medical auxiliary diagnosis of AOSLO.High-accurately AOSLO image registration is also realized via the feature matching algorithm based on nonlinear scale space,and high-order polynomial geometric transformation.With this technology,the multi-frame superposition average of AOSLO image is accomplished,and the image quality is effectively improved.In addition,with the high-accurate image registration technique,continuous tracking and the motion procedure analysis of the ROI of the retina are accomplished,and the averaged tracking error is less than 0.5 pixel.Thus it provides an effective means for the research of the physiological characteristics of the retina,human visual and its working mechanism.The research works mainly include:(1)On the basis of fully understanding the AOSLO imaging system and its core devices,the mathematical model of AOSLO imaging system is given,and the influence of ocular aberration on AOSLO imaging quality is analyzed.Based on the analysis of the distribution and detection environment of residual ocular aberrations,it is pointed out that the reconstruction of residual ocular wavefront aberrations has the problems of anti-noise and local-aberration restoration.To solve this problem,a fast wavefront reconstruction algorithm based on second-order B-spline function is proposed.The local wavefront is fitted by the second-order B-spline function with smoothness and tight support characteristics,which retains the local-aberration restoration ability and has stronger robustness.Experimental results show that the proposed algorithm can effectively improve the accuracy of wavefront restoration,and reduce the residual mean square error of local wavefront reconstruction by about 88%,compared with the traditional method.(2)A series of image processing methods are used to improve the optical imaging quality of AOSLO.Currently,the AOSLO image quality is mainly improved with iterative blind deconvolution algorithms.However,due to the illness nature of the blind restoration problem,these algorithms have poor robustness and easily fall into local optimal solution.To avoid poor robustness of traditional algorithms,this paper firstly proposes a content-adaptive denoising method(CAID)to preprocess degraded AOSLO images.By analyzing simultaneous sparse coding error distributions in different image structures,CAID establishs the optimal denoising problem based on content-adaptive structure group.The experimental results show that the ringing effect of the restoration results is effectively suppressed by using the pre-denoising AOSLO image and traditional blind deconvolution algorithms.However,the problem that the algorithm is easy to fall into the local optimal solution remains to be solved.Aiming at the problem,this paper proposes a blind deconvolution algorithm based on PSF estimation in multiscale image gradient space.In the multi-scale image gradient space,the PSF restoration procedure is regularized by the sparse distribution prior model of intensity and gradient of PSF,and the PSF measured by shack Hartmann wavefront sensor is used as the precision constraint term.After getting the optimal PSF estimation,Wiener filter is used to deconvolute the pre-denoised degraded image and then the final restored image is obtained.In addition,in order to analyze the influence of PSF initial estimation accuracy on the convergence and restoration effect of the proposed algorithm,the traditional Zernike modal method and the second-order B-spline function wavefront reconstruction algorithm proposed in Chapter 2 are used to calculate the initial PSF estimates,which are used as the input of blind deconvolution algorithm.The experimental results show that utilizing the PSF measured by the second-order B-spline wavefront reconstruction method as the initial input of the blind deconvolution algorithm can significantly improve the convergence speed of the algorithm.Based on above algorithm,the images of normal human visual cell layer and nerve fiber layer,as well as images recorded from patients’ eyes are restored.The experimental results show that compared with traditional blind deconvolution algorithms,the proposed scheme recovers more image details and further suppresses noise.Therefore,it is concluded that the proposed method effectively improves the resolution and contrast of AOSLO images,and lays a solid foundation for the medical auxiliary diagnosis of AOSLO.(3)In this paper,the feature algorithm based on nonlinear scale space,and high-order polynomial geometric transformation are used to register AOSLO images.Based on above image registration method,the multi-frame superposition average,highprecision continuous tracking of ROI,and motion analysis of ROI are accomplished.Firstly,by observing the features of AOSLO retinal image in different scale space,it is proposed that nonlinear scale space is more suitable for extracting AOSLO image features.Therefore,the accelerated-KAZE(AKAZE)feature algorithm,which extracts image features from nonlinear scale space,is introduced to find matching feature points of AOSLO images.Once the matching feature points are obtained,the geometric transformation model parameters need to be calculated to register the image.To find the optimal geometric transformation for AOSLO image registration,this paper firstly fits the retinal motion by high-order polynomials,and obtains the corresponding highorder polynomial transformation model.Since the matched feature points of images are limited,the optimal high-order polynomial transformation should satisfy the requirements that the image distortions can be efficiently corrected without introducing additional errors under the limited number of matching points.By observing the registration results of different-order polynomial transformations for multiple datasets,it is suggested that the best polynomial transformation model for AOSLO image registration is the fourth-order polynomial transformation.When registering AOSLO images,due to the influence of inter-frame and intra-frame distortion,the matched feature points close to the image boundary will be missing.Therefore,this paper proposes to optimize the boundary detection range of AKAZE feature algorithm to find more matched feature points close to the image boundary.In addition,an automatic detection and segmentation method of maximum effective region is proposed,which can significantly reduce the image registration error.With above registration method,this paper accomplishs the multi-frame superposition average of AOSLO image.The experimental results show that compared with the traditional algorithm,the proposed method can significantly reduce the image noise and maintain a higher contrast of the average image.Moreover,with the high-accurate image registration technique,continuous tracking and motion procedure analysis of the ROI of the retina are realized,and the averaged tracking error is less than 0.5 pixel,which provides an effective approach to study the physiological characteristics of the retina,the human visual function and its working mechanism.