Research On Image Analysis Of Ophthalmic Optical Coherence Tomography Images

Eye is one of the most important sensory organs of the human being.Visual disorders and eye diseases can bring pains and inconveniences,however,early examination and treatment will be able to help delay pain or cure eye diseases.optical coherence tomography(OCT),especially spectral-domain OCT,appears in the last 25 years,can image anterior segment and poster segment of human eye in a fast,non-contact,non-invasive,high-resolution,and two-dimensional(or threedimensional)way,which offers great conveniences for doctors.Rapid imaging of OCT brings a lot of data.Thus doctors hope to reduce labour work and experience-dependence by way of computer-aided diagnosis.However,OCT image analysis and recognition is lagging seriously behind the rapid development of OCT imaging.In this paper,image segmentation of retina and cornea has been systematically studied.The main contributions include:(1)A customized active contour model is proposed.By exploiting the layered structures of retina and cornea,the traditional active contour model is simplified to include only the image energy and thus decrease computational costs for each layer interface.(2)A customized edge detection for initial location of corneal layer interfaces is proposed.A customized edge detection for initial location of interfaces,fitting the initial interfaces to circles via customized Hough transform is presented.The Hough transform based on the theorem of perpendicular bisection of a chord,is introduced to reduce the computational cost to fit layer interfaces to circles.(3)Kalman filtering to model the correlation between adjacent retinal image frames is proposed.As a prior knowledge,for a volumetric data,the positions of layer interfaces in adjacent images(frames)are similar.The coarse layer interfaces of the current frame could be estimated from the layer interfaces of the previous frame.It has the advantages of avoiding the initialization of contours from the second frame and enhancing the robustness in the presence of retinal blood vessel shades and other artifacts of motion or uneven illumination.(4)Kalman filtering to model horizontal projectile motion of particles for segmenting corneal interfaces is proposed.Through modeling each corneal layer interface as a horizontal projectile motion of particles,Kalman filtering is employed to track edges(particles)in each interface of a single image to yield fast and accurate layer interfaces.Kalman filtering can handle well the heavy noise exhibited in the image and can be adapted to shape variation from a circle to be closer to the real layer interface through a prediction and correction mechanism.Combined with the above approaches,the proposed segmentation frameworks for retina and cornea in this paper were tested for six datasets,including two datasets of normal retinal images,one dataset of retinal images with age-related macular degeneration,one dataset of normal corneal images,one dataset of non-normal corneal images with LASIK flap and one dataset of non-normal corneal images with keratoconus.Experiments show that the proposed method can,as compared with a state-of-the-art method,yield similar or significantly better accuracy and is 37 times and 2.37 times faster for retinal segmentation and cornea segmentation,respectively.