Abstract
Purpose: :
Conventional phenotyping of age-related macular degeneration (AMD) is based on color fundus photographs (CFP). To efficiently extract drusen area and volume measurements from SD OCT imaging, we need automated segmentation of drusen from the images. We compare the areas identified as drusen from semi-automated segmentation of SD OCT to those identified from review of digital CFP. We seek to explore the strengths and limitations of each method of identifying drusen, and to help guide further refinements of our segmentation algorithm.
Methods: :
Ten eyes from ten patients with AREDS Level 3 AMD were imaged with digital CFP and with SD OCT (Bioptigen, Inc. Durham, NC). CFPs were manually marked for drusen within a central zone by three retinal specialists, creating a composite map of drusen which identified areas in which all three graders agreed. Within the same zone, automated drusen segmentation was performed on the SD OCT images, using a novel program which identifies suspect areas based on irregularities in the RPE contour, and then refined by an expert SD OCT reader. The CFP drusen map was compared to an area map of drusen created from the SD OCT analysis (retinal projection). We then studied areas of disagreement to identify the source of discrepancies.
Results: :
Analysis of SD OCT and CFP mapped drusen area produced the following results: They strongly agreed in identification of central areas of intermediate and large drusen. Discrepancies were greatest at drusen borders. On average, SD OCT mapped greater drusen area compared to the CFP map, and SD OCT cross-sections confirmed drusen material extended into sites that were unmarked by experts on CFP. These tended to be in zones of lower elevation of the RPE or increased pigmentation. Other disagreements were due to SD OCT not detecting isolated, small drusen due to sparse sampling, registration errors, and hypopigmented sites in the CFP misidentified as drusen.
Conclusions: :
CFP and SD OCT have demonstrated strong agreement as tools for detecting intermediate and large drusen in these patients with AMD. In numerous cases, SD OCT revealed inaccuracies in CFP analysis by expert graders. Understanding the source of discrepancies between imaging modalities will improve our application of SD OCT in AMD. Because of its potential for greater sensitivity, SD OCT shows great promise for use in a high volume setting for assessing AMD severity.
Keywords: drusen • age-related macular degeneration • imaging/image analysis: clinical