Purchase this article with an account.
Per G Soderberg, Curry Bucht; Fully Automated Frequency Domain Based Estimation Of Cell Densities At Low Cell Count Images Of The Corneal Endothelium. Invest. Ophthalmol. Vis. Sci. 2014;55(13):2044.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Estimation of in vivo corneal endothelial cell density is commonly based on semi-automated analysis of endothelial images, captured by specular microscopy. Due to the need for operator involvement, the semi-automatic estimation of cell density is consuming. The purpose of the presented study was to compare a newly developed fully automated method for estimation of endothelial cell density in images captured by clinical specular microscopy in vivo, to a semi-automated method provided by the IMAGENET-640, using a large population of low cell count images.
The analysis software was developed in MATLAB. Pictures of corneal endothelial cells captured by specular microscopy were read into the software. The pictures were automatically enhanced. Next, the image data was transformed from the spatial domain to the frequency domain using a numerical Fourier transform. Tools for extraction of relevant image characteristics and analysis of frequency domain data were applied to the transformed images. Data obtained from each transformed image was used to estimate the endothelial cell density. Alltogether, 250 in vivo images from 250 patients were analyzed. The endothelial cell density spanned a large range of cell densities. The endothelial cell density estimated using the presently developed method for fully automated analysis was compared to the endothelial cell density estimated by the semi-automated IMAGENET-640 barrier tracing method. Ill conditioned images were excluded from the study.
The average endothelial cell density of the 250 images was 2 678 cells/mm2, spanning [1199; 4750] cells/mm2. Fully automated estimation of endothelial cell density was fitted to semiautomatically estimated endothelial cell density assuming a linear relationship. A 95 % confidence interval for the residual variance was estimated to [183; 218] (cells/mm2)2. corresponding to a coefficient of variation of [0.068; 0.081]. Analysis time for one image was on the order of 15-20 seconds but is possible to reduce with more efficient programming.
Frequency domain estimation of endothelial cell density after primary image enhancement provides reliable estimates of cell countings. The short analysis time compared to semi-automatic estimation is encouraging.
This PDF is available to Subscribers Only