Effective drusen segmentation from fundus images for age-related macular degeneration screening

Abstract

Automatic screening of Age-related Macular Degeneration (AMD) is important for both patients and ophthalmologists. The major sign of contracting AMD at the early stage is the appearance of drusen, which are the accumulation of extracellular material and appear as yellowwhite spots on the retina. In this paper, we propose an effective approach for drusen segmentation towards AMD screening. The major novelty of the proposed approach is that it employs an effective way to train a drusen classifier from a weakly labeled dataset, meaning only the existence of drusen is known but not the exact locations or boundaries. We achieve this by employing Multiple Instance Learning (MIL). Moreover, our proposed approach also tracks the drusen boundaries by using Growcut, with the output of MIL as initial seeds. Experiments on 350 fundus images with 96 of them with AMD demonstrates that our approach outperforms the state-of-the-art methods on the task of early AMD detection and achieves satisfying performance on the task of drusen segmentation.