Purchase this article with an account.
Syouhei Yamada, Naoki Okumura, Mizuki Okamoto, Kaho Hishinuma, Satoru Hiwa, Tomoyuki Hiroyasu, Noriko Koizumi; Development of a deep neural network system for the analysis of specular microscopy images. Invest. Ophthalmol. Vis. Sci. 2020;61(7):1441.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
The analysis of corneal endothelium via specular microscopy is important for the clinical evaluation of patients with Fuchs corneal endothelial dystrophy (FECD). However, making a proper objective assessment is challenging due to the presence of guttae. The purpose of this present study was to develop an artificial neural network by deep learning for the analysis of corneal endothelium and guttae in FECD.
In this study, mutant knock-in mice with Col8a2 (Q455K/Q455K) (Jun AS, et al. Hum Mol Genet. 2012) were used as an FECD mouse model. Images of the corneal endothelium in the FECD mouse model were obtained via contact specular microscopy. Correct-answer data was prepared from 20 images via manual illustration of the cells and guttae margins, with that data then being divided into five subsets (n = four images each). Four subsets were used as training data, and the one subset were used as test data. The training/testing was performed to produce a learned model via application of the U-Net (convolutional network architecture for fast and precise segmentation of images). Pearson’s correlation coefficient was used to determine correlations between the correct answer and the data provided by learned model for the correlation of cell density (CD), coefficient of the variation (CV) and cell hexagonality (6A).
A learned model was developed by deep learning using U-Net, and it successfully visualized the margin of the cells and guttae of the FECD mouse model images obtained via contact specular microscopy. Pearson’s correlation coefficient showed a very strong correlation between CD determined by the learned model and the correct answer of CD (r=0.947, p<0.001). Likewise, CV and 6A determined by the learned model showed a very strong correlation with those correct answers (r=0.817 and r=0.867, respectively; p<0.001). Rates of concordance between guttae determined by the learned model and correct answer of guttae was 99.1%.
The findings in this study involving the use of FECD mouse-model images shows that a deep neural network is applicable for the analysis of corneal endothelial cells and guttae, and similar to these current findings, we anticipate the development of a deep neural network applicable for human data.
This is a 2020 ARVO Annual Meeting abstract.
This PDF is available to Subscribers Only