Purchase this article with an account.
Mengyu Wang, Lucy Q Shen, Louis R Pasquale, Thao D. Nguyen, Yangjiani Li, Mohammad Eslami, Chhavi Saini, Nazlee Zebardast, Tobias Elze; Artificial Intelligence Assessment of Optic Cup Shape Patterns in Glaucoma. Invest. Ophthalmol. Vis. Sci. 2021;62(8):3359.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To assess optic cup shape patterns in glaucoma with unsupervised artificial intelligence (AI).
The first Cirrus OCT scans of the optic nerve head (ONH) from each eye with signal strength ≥ 6 were selected. The optic cup shape was represented as the vertical positions of the inner limiting membrane (ILM) with respect to the lowest ILM vertical position in each eye. Scans with ONH centers deviating more than 0.3 mm from the scan center were excluded. The OCT scans were registered with respect to ONH centers in the right eye format. The Humphrey SITA standard 24-2 visual fields (VFs) tested within three months of the OCT tests were selected. An unsupervised AI method termed non-negative matrix factorization was applied to assess the cup shape patterns. The cup shape patterns were correlated with VF and OCT diagnostic parameters. We compared if using the cup shape patterns improved the prediction of VF loss using linear regression with model selection to remove redundant features.
We determined 14 cup shape patterns (Figure 1) from 9,854 OCT scans. The brighter regions in the cup shape patterns indicate the more informative zones with greater variations across patients. Mean deviation (MD), retinal nerve fiber layer thickness (RNFLT) and ONH related parameters including rim area, disc area, average cup-disc (CD) ratio, vertical CD ratio and cup volume (Table 1) were most negatively correlated with Patterns 5, 4, 4, 12, 14, 14 and 14 (r: -0.15, -0.21, -0.45, -0.42, -0.60, -0.57 and -0.49, p < 0.001), and were most positively correlated with Patterns 10, 10, 14, 5, 5, 7 and 5 (r: 0.30, 0.28, 0.39, 0.54, 0.59, 0.52 and 0.78, p < 0.001), respectively. The Worse MD and thinner RNFLT were most strongly correlated with higher coefficients of Patterns 10 and 12, which represent inferior and superior cupping. The adjusted multiple r (rm) to predict MD separately using the cup shape patterns, ONH related parameters, and 12 clock hour RNFLTs were 0.50, 0.51 and 0.56, respectively, which the model combining the cup shape patterns and ONH related parameters (adjusted rm: 0.60, p < 0.001) outperformed. The model (adjusted rm: 0.63) combining all three types of features outperformed (p < 0.001) the model combining existing RNFLTs and ONH parameters (adjusted rm: 0.59).
The cup shape patterns correlated with established diagnostic parameters and improved the structure-function relationship in glaucoma.
This is a 2021 ARVO Annual Meeting abstract.
This PDF is available to Subscribers Only