Research Of Registration Algorithms For 3D Retinal OCT Images

Medical image registration has attracted more and more attention due to its valuable applications in clinical studies.In particular,registration can be used for studying longitudinal and cross-sectional data,quantitatively monitoring disease progression and guiding computer assisted diagnosis and treatments.Registration is a fundamental task in medical image processing used to match multiple images taken from different viewpoints,different sensors or different time points.In the past decades,many registration techniques have been developed for various types of data and applications.However,spectral-domain optical coherence tomography(SD-OCT)image registration which enables more precise and quantitative comparison of retinal disease has not been well developed.3D SD-OCT imaging technique is a noninvasive and non-contact scan of the retina,which became available in 2007 and has been widely used in investigation of retinal pathology,such as Choroidal Neovascularization(CNV),age-related macular degeneration,glaucoma and so on.Since SD-OCT imaging is relatively new compared with other medical imaging modalities such as magnetic resonance imaging and computed tomography,the requirement for processing SD-OCT images has a shorter history.With the fast development of SD-OCT technique,the demand for advanced image analysis techniques is rapidly growing.Nevertheless,the development of such techniques can be challenging as SD-OCT image is inherently noisy and the structure of retina can change drastically when pathology occurs.Therefore,although many image registration techniques have been well developed,they cannot be directly applied to SD-OCT images.In this thesis,the registration method of 3D SD-OCT imaging of retina is studied,and registration methods for both normal retina and severely damaged retina are proposed.The main contributions of this thesis are summarized as follows:(1)The preprocessing methods of OCT image registration are proposed,which include retinal layer segmentation algorithm and eye motion correction algorithm.In recent years,graph cuts and graph search techniques are two graph based segmentation approaches which are successfully applied to 3D medical image segmentation.The graph cut method can find the global optimal solution for many energy minimization problems,which is suitable for the segmentation of regions of interest,while the graph search method is particularly suitable for retinal layer segmentation problem.Considering the multiple surfaces structure of the retina and the characteristics of the lesion area,a multiple surfaces graph search approach is applied to segment the retinal layers.According to the layer segmentation result,a B-scan flattening method is proposed to eliminate the axial eye movement.(2)A vascular feature-based registration method for OCT projection image is proposed.Serious speckle noise of OCT results in poor signal to noise ratio(SNR)compared with other medical imaging modalities.The normal intensity based registration methods tend to be more sensitive to the noises during the registration process.For OCT projection image registration,the vessel maps of OCT projection images between the template and the subject are registered to find out x-y direction displacement.In order to improve the quality of OCT projection image,firstly,denoising and enhancement processing are carried out.The multi-scale vessel enhancement filter and morphological thinning methods are then used to extract the vessel maps from the projection image of 3D OCT scans.And finally,the x-y direction displacement is estimated by matching SpeededUp Robust Features of the vessel maps.(3)An OCT image registration algorithm based on A-scan aligment is proposed.Since a single A-scan is acquired at a time,voxels have strong correlations along A-scans.Therefore,each A-scan is considered as a base deformable unit and the voxel transformation is not allowed across different A-scans.Therefore,the mis-matching across different A-scans is avoided.Considering the serious speckle noise,the tissue map is used instead of the intensity image in the registration to reduce the negative effect of the speckle noises.Therefore,the registration accuracy is improved.(4)A full 3D deformable registration approach for retinal OCT images which can be applied to longitudinal OCT images for both normal and serious pathological subjects is proposed.The deformable transformation is a free form mapping at each voxel x.It can be solved by finding a transformation of each voxel such that the energy function is minimized.Considering the high resolution of OCT data,the energy function would be a very high dimension function which makes it extremely difficult to find the global optimal solution.The main difficulties are the computation complexity and the local minima problem.To speed up registration process and reduce local minima,a novel designdetection-deformation mechanism is designed.According to the segmented 7 retinal surfaces,intensity-based region feature,surface-based structure feature and vessel-like feature are designed for the registration.To speed up registration process and reduce local minima,a hierarchical strategy is also applied.Active voxels are hierarchically selected and the point-to-point correspondences between the subject and the template images are established.The image is then hierarchically deformed according to the detected correspondences in multi-resolution.This study enriches and improves the non-rigid registration theory,and forms rapid and accurate OCT registration software,which is helpful to assist ophthalmologists to better analyze the progress of treatment and make treatment plans.