Purchase this article with an account.
Satish Kumar Panda, Tin A. Tun, Shamira Perera, Ching-Yu Cheng, Tin Aung, Alexandre Thiéry, Michael J A Girard; Use of Artificial Intelligence to Describe the Structural Signature of the Glaucomatous Optic Nerve Head. Invest. Ophthalmol. Vis. Sci. 2021;62(8):1030.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To investigate the complex structural changes in the optic nerve head (ONH) with the development of glaucoma and to find novel biomarkers to describe the structural phenotype of the glaucomatous ONH.
We recruited 3,782 subjects (2233 glaucoma, 1549 non-glaucoma) and imaged the ONH of each subject using optical coherence tomography (OCT). Using a deep learning network, we automatically segmented seven neural and connective tissue layers of the ONH. The segmented OCT images were then reduced to a few latent features (LFs) and reconstructed using an autoencoder. A network with MLP layers was used concurrently to classify the images as glaucoma or non-glaucoma from the LFs. We then performed a principal component analysis (PCA) on LFs and identified the principal components (PCs). Subsequently, we altered the magnitude of each PC in steps and reported how it impacted the morphology of the ONH. To facilitate visualization, we used Uniform Manifold Approximation and Projection (UMAP) to further reduce the LFs to a 2D space.
The image reconstruction quality and diagnostic accuracy increased with the numbers of LFs; with 54 LFs, the Dice coefficient for the reconstructed images was 0.86±0.04, and the accuracy was 92.0±2.3%. Using UMAP, we were able to dissociate non-glaucoma eyes from glaucoma eyes into two distinct clusters (Fig. 1; AUC = 0.96). The PC1 values for non-glaucoma and glaucoma eyes were significantly different (p<10−9), and by varying it from a high value (non-glaucoma eye) to a low value (glaucoma eye), we observed thinning of neural and connective tissue layers, decrease in minimum rim width, and increase in prelamina depth. When PC1 was changed incrementally, we observed ONH structural changes (non-glaucoma to glaucoma zone) that may reflect progression.
Our network identified novel biomarkers and revealed the structural changes in the glaucomatous ONH. PC1 summarized multiple and simultaneous structural variations into a single number, and by changing its magnitude, the ONHs transitioned from a non-glaucoma to a glaucoma zone. The paradigm introduced herein may help us describe the structural phenotype of the glaucomatous ONH.
This is a 2021 ARVO Annual Meeting abstract.
Left: The structural changes of ONH during glaucoma development. Right: ONH transition from the non-glaucoma (blue point cloud) to glaucoma (redpoint cloud) zone while varying PC1 for a given non-glaucoma ONH starting point.
This PDF is available to Subscribers Only