Fast, Accurate and Automatic Segmentation and Classification of Ophthalmic Optical Coherence Tomography Images Based on Sparse Representation

点击翻译按钮获取中文摘要

The research program proposed in this application aims at fast & accurate classification of ophthalmic Optical Coherence Tomography (OCT) images based on sparse representation. Manual analysis of ophthalmic images is a time and labor intensive subjective procedure. Therefore, in recent years automatic analysis of such images has gained a lot of interest. One of the more recent modalities used to obtain detailed images from within the retina is OCT. OCT is one of the fastest adopted technologies in ophthalmology for diagnosis and study of retinal pathologies. Combining OCT technologies with image processing techniques provides detail information about different internal layers of retina that are crucial for diagnosis of diseases such as glaucoma, diabetic retinopathy, and age-related macular degeneration (AMD) - the three leading causes of blindness. Although a fair amount of interesting work has been done to develop fully/semi- automatic OCT image analysis, these attempts have been usually limited to preprocessing or segmentation of intra-retinal layers. The main goal of this research program is to introduce a novel modeling approach that allows automatic classification of OCT data into 4 categories: 1) normal, 2) glaucoma, 3) diabetic retinopathy, and 4) AMD. The main advantage of automatic classification of OCT data is in its application in remote monitoring devices. Remote monitoring will reduce doctor visits by eliminating unnecessary referrals which in turn improves the care provided to those who actually need special care. To reach the main goal of this research, the applicant proposes an atomic representation modeling. At first, the best atoms (basis functions of our transform, also known as dictionary) are found that fit best on each category. These atoms can produce a new model for the OCT data. The coefficients of the atomic representation can be used for classification purpose. For example, for normal data each intra-retinal layer can be represented by its specific independent atoms; therefore, the normal data can be reconstructed by a combination of these atoms. Similarly, a different set of atoms (and so a different model) for AMD data are investigated, i.e., the proposed atoms for cysts in AMD are different from each layer’s atoms. After finding the best model for each category, the obtained (sparse) coefficients are used for classification instead of directly classifying the images.