Purchase this article with an account.
Tristan Hormel, JIE WANG, Qisheng You, David Huang, Thomas Hwang, Yali Jia; Automated Detection of Retinal Fluid Using a Convolutional Neural Network. Invest. Ophthalmol. Vis. Sci. 2019;60(11):PB087.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Diabetic macular edema (DME) is a major cause of vision loss and is not amenable to screening with photographic methods. Retinal fluid (RF) on optical coherence tomography (OCT) may be an earlier sign than macular thickening to identify eyes at risk for vision loss from DME. We developed a convolutional neural network (CNN) for automated detection of RF in OCT volumes.
A commercial spectral-domain OCT system (RTVue-XR, Optovue, Inc.) captured OCT macular scans (304×304 A-lines covering 3×3 or 6×6 mm2 area) from both eyes of a total 172 patients with diabetes and generated 2150 scan volumes by including repeat visits. An expert graded each volume for the presence of fluid. We reserved 200 volumes for testing and used the remainder for training. We did not exclude any scans due to poor image quality.Our CNN architecture consists of data compression, feature extraction and decision-making steps (Fig. 1). Data compression was accomplished through successive convolution, batch normalization and rectified linear unit activation. Feature extraction was achieved by additionally including max pooling layers. Finally, fully connected layers and a softmax activation function made the prediction. Volumes with a probability over 50% of containing RF were classified as positive for its presence.
Compared to expert manual grading, the trained network classified the volumes for the presence of RF with an accuracy of 83.33%, a precision of 84.62%, and recall of 81.05% on the test data set. Fig. 2 shows example B-scans with the associated detection probability.
Our CNN detects RF with high accuracy on a diverse data set regardless of scan quality or severity of disease. Volumetric evaluation obviates the need for selection of specific B-scans. Inclusion of both 3×3 and 6×6 mm2 scans makes our results agnostic to scan dimensions.
This abstract was presented at the 2019 ARVO Imaging in the Eye Conference, held in Vancouver, Canada, April 26-27, 2019.
This PDF is available to Subscribers Only