Peer Reviewed

1

DOI

10.17077/omia.1026

Conference Location

Munich, Germany

Publication Date

October 2015

Abstract

Age-related macular degeneration (AMD) is a leading cause of blindness in developed countries. Its most damaging form is characterized by accumulation of fluid inside the retina, whose quantification is of utmost importance for evaluating the disease progression. In this paper we propose an automated method for retinal fluid segmentation from 3D images acquired with optical coherence tomography (OCT). It combines a machine learning approach with an effective segmentation framework based on geodesic graph cut. After an image preprocessing step, an artificial neural network is trained based on textural features to assign to each voxel a probability of belonging to a fluid. The obtained probability maps are used to compute minimal geodesic distances from a set of identified seed points to the remaining unassigned voxels. Finally, the segmentation is solved optimally and efficiently using graph cut optimization. The method is evaluated on a clinical longitudinal dataset consisting of 30 OCT scans from 10 patients taken at 3 different stages of treatment. Manual annotations from two retinal specialists were taken as the gold standard. The segmentation method achieved mean precision of 0.88 and recall of 0.83, with the combined F1 score of 0.85. The segmented fluid volumes were within the measured inter-observer variability. The results demonstrate that the proposed method is a promising step towards accurate quantification of retinal fluid.

Rights

Copyright © 2015 Hrvoje Bogunović, Michael D. Abràmoff, and Milan Sonka

Included in

Ophthalmology Commons

Share

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

Geodesic Graph Cut Based Retinal Fluid Segmentation in Optical Coherence Tomography

Munich, Germany

Age-related macular degeneration (AMD) is a leading cause of blindness in developed countries. Its most damaging form is characterized by accumulation of fluid inside the retina, whose quantification is of utmost importance for evaluating the disease progression. In this paper we propose an automated method for retinal fluid segmentation from 3D images acquired with optical coherence tomography (OCT). It combines a machine learning approach with an effective segmentation framework based on geodesic graph cut. After an image preprocessing step, an artificial neural network is trained based on textural features to assign to each voxel a probability of belonging to a fluid. The obtained probability maps are used to compute minimal geodesic distances from a set of identified seed points to the remaining unassigned voxels. Finally, the segmentation is solved optimally and efficiently using graph cut optimization. The method is evaluated on a clinical longitudinal dataset consisting of 30 OCT scans from 10 patients taken at 3 different stages of treatment. Manual annotations from two retinal specialists were taken as the gold standard. The segmentation method achieved mean precision of 0.88 and recall of 0.83, with the combined F1 score of 0.85. The segmented fluid volumes were within the measured inter-observer variability. The results demonstrate that the proposed method is a promising step towards accurate quantification of retinal fluid.