Fast detection and segmentation of drusen in retinal optical coherence tomography images

Accurate detection-characterization of drusen is an important imaging biomarker of age-related macular degeneration (AMD) progression. We report on the development of an automatic method for detection and segmentation of drusen in retinal images captured via high speed spectral domain optical coherence tomography (SDOCT) systems. The proposed algorithm takes advantage of a priori knowledge about the retina shape and structure in the AMD and normal eyes. In the first step, the location of the retinal nerve fiber layer (RNFL) is estimated by searching for the locally connected segments with high radiometric vertical gradients appearing in the upper section of the SDOCT scans. The high reflective and locally connected pixels that are spatially located below the RNFL layer are taken as the initial estimate of the retinal pigment epithelium (RPE) layer location. Such rough estimates are smoothed and improved by using a slightly modified implementation of the Xu-Prince gradient vector flow based deformable snake method. Further steps, including a two-pass scan of the image, remove outliers and improve the accuracy of the estimates. Unlike healthy eyes commonly exhibiting a convex RPE shape, the shape of the RPE layer in AMD eyes might include abnormalities due to the presence of drusen. Therefore, by enforcing local convexity condition and fitting second or fourth order polynomials to the possibly unhealthy (abnormal) RPE curve, the healthy (normal) shape of the RPE layer is estimated. The area between the estimated normal and the segmented RPE outlines is marked as possible drusen location. Moreover, fine-tuning steps are incorporated to improve the accuracy of the proposed technique. All methods are implemented in a graphical user interface (GUI) software package based on MATLAB platform. Minor errors in estimating drusen volume can be easily manually corrected using the user-friendly software interface and the program is constantly refined to correct for the repeating errors. This semi-supervised approach significantly reduces the time and resources needed to conduct a large-scale AMD study. The computational complexity of the core automated segmentation technique is attractive as it only takes about 6.5 seconds on a conventional PC to segment, display, and record drusen locations in an image of size (512 × 1000) pixels. Experimental results on segmenting drusen in SDOCT images of different subjects are included, which attest to the effectiveness of the proposed technique.

Duke Scholars

Author Sina Farsiu Biomedical Engineering

Author Cynthia Ann Toth Ophthalmology, Vitreoretinal Diseases & Surgery