Purpose

Age-related macular degeneration (AMD) is a progressive eye condition that results in loss of central vision and severely impacts quality of life for over 15 million elderly Americans. Grading of primary form of AMD (dry AMD) is done by quantifying area of drusen bodies, a process that is slow and error prone when done manually. To address this, we present a novel drusen detection and quantification scheme that is an essential first step to a fully-automated AMD screening system that can be applied to flag early signs of AMD.

Methods

Our technique combines robust low-level image processing and powerful statistical inference to enable fully-automated drusen identification from color fundus images. The steps include image normalization, region of interest detection, detected pixel description, pixel level classification, and blob level description and classification. We evaluate the method on a set of 40 images that were drusen annotated by experts at Doheny Eye Institute (DEI) in a region centered at the fovea with a radius of twice the ONH diameter. At each pixel the system evaluates low-level image properties (local edges, textures, brightness, contrast, color etc.) in a multi-scale framework, composes them into descriptors, classifies them using advanced machine learning techniques and aggregates the results.

Results

Fig. 1 (see link) presents an example of the detected drusen along with the drusen probability indicated by gray scale values. Blob level area under receiver operating characteristic curve (AUROC) for drusen identification on the entire set of images was evaluated to be 0.90 (ROC curve see link).

Conclusions

We present a novel approach for drusen detection that achieves an AUROC of 0.90 on the test dataset and produces visualization with drusen probabilities that can be used for patient education. This would be a key component in a fully automated screening system that would aid early detection and treatment of AMD.