Abstract
Purpose: :
To search for imaging biomarkers of age-related macular degeneration (AMD) progression. Accurate detection of drusen volume and characteristics is important as drusen size and type predicts progression. In previous small-scale SDOCT studies, drusen segmentation and characterization were done manually. In a large-scale multi-year study, with hundreds of patients each with hundreds of scans, a manual approach is impractical. Automatic image analysis is critical to expedite drusen analysis.
Methods: :
Eyes with Age-Related Eye Disease Study Level 3 AMD were imaged, with SDOCT (Bioptigen, Inc. Durham, NC) to create 3D representations of the retina and drusen. Initial automated segmentations of retinal nerve fiber layer followed by retinal pigment epithelium (RPE) layer were further improved by using an iterative deformable snake, which was later corrected for outliers. By fitting locally convex curves to this possibly unhealthy RPE curve, the healthy shape of the RPE layer was estimated. Area between estimated unhealthy and healthy RPE outlines was marked as drusen location. Individual drusen characteristics (e.g. shape, internal reflectivity) previously only evaluated visually, were mathematically defined and evaluated. All methods were implemented in a graphical user interface software package based on MATLAB platform. Total drusen volume and characteristics were measured for the central subfield by the automatic system and then refined with grader-adjusted analysis (GAA).
Results: :
Using this software, resulting volume and characteristics of drusen produced non-overlapping subcategories for eyes with Level 3 AMD. Although most drusen were detected by the software resulting in GAA of less than 25% of volume (compared to interobserver difference after GAA of ~7%); there was a consistent increase in drusen volume with GAA, due to the ad-hoc definitions of the level of the outline defining the base of drusen areas. Automated drusen volume measurement (6 sec per frame), even with GAA, was many times faster than manual grading. Confluent drusen were frequently found by software and verified on GAA in eyes with soft indistinct drusen.
Conclusions: :
The developed software accurately segments and categorizes drusen in the SDOCT images of AMD eyes. Errors can be manually corrected using a 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.
Keywords: age-related macular degeneration • imaging/image analysis: clinical • drusen