SVM Based Retinal IS/OS Disruption Detection From OCT Images

Nearly80%information gained in daily life is from vision. With the change of themodern lifestyle and rapid growth of global aging population, more and more people aresuffering from various retinal diseases. Optical coherence tomography (OCT), ahigh-resolution, noninvasive imaging method, helps the ophthalmologist to view retinalmorphology more clearly and improve the diagnosis. But the problem is the hugeinformation needing to be reviewed. If the images can be preprocessed according toclinical demands, it will greatly improve their working efficiency.Inner segment/outer segment (IS/OS) is an important area consisting ofphotoreceptors, which is essential for vision. Many kinds of diseases will affect theintegrity of IS/OS, which is something related to human vision. This paper proposes amethod to detect the IS/OS disruption region automatically from spectral domain OCT(SD-OCT) images. It can be divided into three parts: preprocessing, training and testing,result showing and analysis. Firstly, the11surfaces for each image are segmented, and theground truth for each image is manually labeled. Secondly, features extracting is applied toOCT images with a total feature number of28, such as normalized intensity, absoluteintensity difference in13directions, block mean value and so on. In training phase,training data sets are completed with a random5:1sample rate. The leave-one-out strategyis used to test the testing data set one by one.The main innovation is that SVM is utilized to classify the disruption andnon-disruption region of the fovea IS/OS region,1mm diameter area around the fovea, andcalculate the accuracy separately. The results demonstrated the feasibility and efficiency ofthe proposed method, with a mean accuracy of82.66%.