Purchase this article with an account.
Keunheung Park, Jiwoong Lee; The relationship between Bruch’s membrane opening-minimum rim width and retinal nerve fiber layer thickness and their combinational index using artificial neural network. Invest. Ophthalmol. Vis. Sci. 2018;59(9):4077.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To evaluate the relationship between Bruch’s membrane opening minimum rim width (BMO-MRW) and peripapillary retinal nerve fiber layer thickness (pRNFLT) and to make a new single parameter combining BMO-MRW and pRNFLT using artificial neural network to maximize their probable compensatory diagnostic performance.
A total of 402 subjects were included and divided into 2 major groups, 273 for validation group and 129 for neural net training. Validation group consisted of: 141 healthy, 70 early glaucoma and 62 advanced glaucoma patients. Both BMO-MRW and pRNFLT were measured with Spectralis optical coherence tomography simultaneously. One way ANOVA test was used to compare BMO-MRW and pRNFLT among 3 patients groups. Linear, quadratic and broken-stick regression model was used to find relationship between BMO-MRW and pRNFLT. Multilayer neural network with 1 hidden layer was used to make single combined parameter and the diagnostic performance was compared using area under receiver operating curve (AUROC).
In regression analysis, the best fitted statistical model between BMO-MRW and pRNFLT was broken-stick model and the worst fitted model was linear regression model. Globally, the significant tipping point was located at 226.5 µm of BMO-MRW. Both BMO-MRW and pRNFLT was significantly correlated with visual field mean deviation (MD) and the best fitted model was the broken-stick model. When comparing diagnostic power between BMO-MRW and pRNFLT, BMO-MRW showed higher AUROC than pRNFLT. In differentiating normal and glaucoma, neural network showed significantly the largest AUROC compared to both BMO-MRW and pRNFLT. In differentiating normal and early glaucoma, overall diagnostic performance was decreased in all of the three parameters but among them, neural network still showed significantly the best AUROC.
The relationship between BMO-MRW and pRNFLT showed very significant broken-stick relationship. Before pRNFLT starts thinning, the considerable BMO-MRW thinning was preceded. The neural network largely improved the diagnostic power by combining both BMO-MRW and pRNFLT.
This is an abstract that was submitted for the 2018 ARVO Annual Meeting, held in Honolulu, Hawaii, April 29 - May 3, 2018.
This PDF is available to Subscribers Only