Abstract
Purpose :
To develop a deep learning algorithm that can provide a glaucoma diagnosis from a single optical coherence tomography (OCT) scan of the optic nerve head (ONH) by exploiting 3D structural information about neural and connective tissues.
Methods :
A diagonal OCT scan (6 B-Scans; Spectralis) was acquired through the center of the ONH for 1 eye of each of 2,701 Chinese subjects from the Beijing Eye Study (2,566 subjects were clinically classified as non-glaucoma, and 135 as glaucoma). We developed a custom convolutional neural network (deep learning) to: 1) accurately segment important neural and connective tissue structures of the ONH (retinal nerve fiber layer thickness [RNFL], prelamina, retinal pigment epithelium, all other retina layers, lamina cribrosa, choroid, and peripapillary sclera); and 2) to combine segmentation information with OCT images to classify glaucoma from non-glaucoma subjects. For segmentation training, we used 10 manually-segmented images, and for glaucoma diagnosis training we used 70% of the images (from 1,891 subjects). The rest of the images (810 subjects) were used for diagnosis testing. We also performed extensive data augmentation (e.g. image deformations) to improve ONH segmentation accuracy. We calculated the area under the receiver operating characteristic curve (AUC) to assess whether our deep learning algorithm was able to differentiate glaucoma from non-glaucoma eyes. Finally, we also compared our diagnostic performance with that of RNFL thickness (measured from circular scans).
Results :
Our deep learning algorithm was able to accurately segment neural and connective tissues of the ONH with a performance almost as good as that of a human (Figure 1A). In addition, our algorithm was able to classify glaucoma from non-glaucoma with an AUC of 0.90 (Figure 1B). This value was higher than those derived from RNFL thickness alone: 0.79 (Infero-temporal), 0.70 (infero-nasal), 0.77 (supero-nasal), and 0.70 (supero-temporal).
Conclusions :
We propose a novel deep learning approach that can exploit 3D structural information of ONH neural and connective tissues to provide a glaucoma diagnosis from a single OCT scan. Our approach may simplify glaucoma diagnosis as it is fast (diagnosis is computed in less than 1 second), and it only requires 1 clinical test (i.e. OCT imaging of the ONH).
This is an abstract that was submitted for the 2018 ARVO Annual Meeting, held in Honolulu, Hawaii, April 29 - May 3, 2018.