Peer Reviewed

1

DOI

10.17077/omia.1029

Conference Location

Munich, Germany

Publication Date

October 2015

Abstract

Age-related Macular Degeneration (AMD) is one of the leading causes of blindness. Automatic screening of AMD has attracted much research effort in recent years because it brings benefits to both patients and ophthalmologists. Drusen is an important clinical indicator for AMD in its early stage. Accurately detecting and localizing drusen are important for AMD detection and grading. In this paper, we propose an effective approach to localize drusen in fundus images. This approach trains a drusen classifier from a weakly labeled dataset, i.e., only the existence of drusen is known but not the exact locations or boundaries, by employing Multiple Instance Learning (MIL). Specifically, considering the sparsity of drusen in fundus images, we employ sparse Multiple Instance Learning to obtain better performance compared with classical MIL. Experiments on 350 fundus images with 96 having AMD demonstrates that on the task of AMD detection, multiple instance learning, both classical and sparse versions, achieve comparable performance compared with fully supervised SVM. On the task of drusen localization, sparse MIL outperforms MIL significantly.

Rights

Copyright © 2015 Huiying Liu, Yanwu Xu, Damon W.K. Wong, and Jiang Liu

Included in

Ophthalmology Commons

Share

COinS
 
Oct 9th, 12:00 AM Oct 9th, 12:00 AM

Effective Drusen Localization for Early AMD Screening using Sparse Multiple Instance Learning

Munich, Germany

Age-related Macular Degeneration (AMD) is one of the leading causes of blindness. Automatic screening of AMD has attracted much research effort in recent years because it brings benefits to both patients and ophthalmologists. Drusen is an important clinical indicator for AMD in its early stage. Accurately detecting and localizing drusen are important for AMD detection and grading. In this paper, we propose an effective approach to localize drusen in fundus images. This approach trains a drusen classifier from a weakly labeled dataset, i.e., only the existence of drusen is known but not the exact locations or boundaries, by employing Multiple Instance Learning (MIL). Specifically, considering the sparsity of drusen in fundus images, we employ sparse Multiple Instance Learning to obtain better performance compared with classical MIL. Experiments on 350 fundus images with 96 having AMD demonstrates that on the task of AMD detection, multiple instance learning, both classical and sparse versions, achieve comparable performance compared with fully supervised SVM. On the task of drusen localization, sparse MIL outperforms MIL significantly.