Purchase this article with an account.
Warren Lewis, Sophie Kubach, Luís de Sisternes, Zhongdi Chu; Optimized segmentation of the choriocapillaris in ophthalmic optical coherence tomography angiography (OCTA) scans via generative adversarial networks. Invest. Ophthalmol. Vis. Sci. 2020;61(9):PB0074.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Segmentation of the choriocapillaris (CC) can be challenging for automated OCTA layer segmentation algorithms because it is relatively thin and hard to identify. It is usually segmented using an offset from other, more easily segmented features, such as the retinal pigment epithelium or Bruch’s membrane. Even though they are more easily segmented, these other layers are also subject to errors. In addition to this, the distance to the CC from these layers is variable. This work is an exploration of the use of deep learning to recognize the appearance of the CC and thereby assist in the segmentation of this layer.
A generative adversarial network (GAN) was trained to produce and discriminate patches of healthy CC slabs. As ground truth, CC en face projections were created from angiography scans of healthy patients with a PLEX® Elite 9000 SS-OCT system (ZEISS, Dublin, CA). The slab boundaries were manually edited to ensure that the CC was accurately located. These slabs were used to create patches for use as ground truth. After training, the output of the Generator component of the GAN was as shown in Figure 1a. For comparison, a sample of patches from the actual CC slabs is shown in Figure 1b. At this point, the Discriminator from this trained network could be used to calculate a CC similarity figure for patches from the test volume.This Discriminator was used to optimize the boundaries of the CC slab in a new test volume. Each point in a grid was used as the center of a patch of the test slab. The point of best similarity to CC was identified at each grid location. This point was used to define the boundary of the CC. The grid of optimized depth locations was then used to interpolate the segmentation surface.
Figure 2a shows an inaccurate CC slab projected with inaccurate segmentation. Figure 2b shows the slab after correction by this algorithm. The algorithm has improved the accuracy of the CC slab boundaries.
The use of deep learning to recognize the appearance of structures in OCTA imaging can improve the segmentation of the CC layer as output by automated segmentation algorithms.
This is a 2020 Imaging in the Eye Conference abstract.
Figure 1. Left, patches taken from actual CC slabs. Right, sample output of the Generator network
Figure 2. Left, incorrectly segmented CC slab. Right, CC slab calculated after correction by the algorithm.
This PDF is available to Subscribers Only