Graph-Based Segmentation for Diabetic Macular Edema Selection in OCT Images

Abstract

Diabetic macular edema results in severe complications leading to blindness and is characterized by specific areas in the optical coherent tomography images (OCT). We propose a technique for diabetic macular edema selection, which is based on the pre-processing of OCT images using the edge detection method and graph-based image segmentation. In the course of study, the value of $\sigma=3.5$ was demonstrated to be an optimal value of the $\sigma$ parameter of a filter kernel utilized at a preprocessing stage. The image binarization threshold in the Canny algorithm was chosen based on a criterion of reduction of spurious edges in the resulting image. The best result was attained at a threshold of 0.6. It has been experimentally demonstrated that when the percentage of minimum cluster size equals 2.5% it is possible to attain a retinal segmentation error of 2%.