Abstract
Abstract: :
Purpose:To demonstrate a new segmentation algorithm for the commercially available optical coherence tomography system. Our algorithm can automatically map several retinal features, including retinal boundaries. This study intends to show that the retinal boundaries can be automatically located with great accuracy even in eyes presenting abnormal/unusual anatomy and/or poor quality images. Methods:STRATUS OCT images are processed using a new noise reduction technique developed specifically for OCT applications. The Structural Correlation Algorithm for Speckle Removal (SCASR) is a non–linear anisotropic filter that greatly reduces the noise variance of OCT images. Unlike standard smoothing filters SCASR does not reduce resolution. Once the images have been cleaned they are analyzed using a new iterative boundary detection algorithm. The algorithm is able to locate automatically and/or interactively several retinal features, including the global boundaries (inner limiting membrane and RPE), the boundaries of the major anatomical layers internal to the retina, boundaries of epiretinal membranes and other vitreous structures, the boundaries of cystic spaces, drusen, and RPE detachment. Results:All OCT images (sets of 6 macular radial scans) taken at the Bascom Palmer Eye Institute retinal clinics for regular patient care over the course of five consecutive days were reviewed. The commercially available retinal thickness algorithm was deemed to have a major failure if the estimated retinal boundaries diverged by more than 50 µm over a distance of at least 20 A–scans from boundaries determined by visual inspection. It was determined that major failures affected at least one scan in approximately 20% of our patients. These were most commonly eyes with complex pathologies (like macular edema, AMD, and vitreoretinal interface disorders) however they also included some retinas whose appearance was substantially normal. 40 B–scans were randomly chosen from the set of images that reported major failures. They were analyzed with our software and the correct retinal boundaries (ILM and RPE) were mapped on 37 out of the 40 selected images. Conclusions:The variety and complexity of retinal anatomy is a challenge for automated segmentation. The presently available algorithm computes incorrect retinal thickness values for a sizeable fraction of our patient population. Our new segmentation algorithm, however, is successful even on very complicated and/or poor quality images . This demonstrates that accurate retinal thickness maps can be obtained from STRATUS OCT images of patients with serious retinal pathologies.
Keywords: imaging/image analysis: clinical • imaging methods (CT, FA, ICG, MRI, OCT, RTA, SLO, ultrasound) • macula/fovea