Purchase this article with an account.
Sa Xiao, Felicitas Bucher, yue wu, Ariel Rokem, Cecilia S Lee, Kyle V Marra, Regis Fallon, Sophia Diaz-Aguilar, Edith Aguilar, Martin Friedlander, Aaron Y. Lee; Web based, fully automated, deep learning segmentation of oxygen induced retinopathy. Invest. Ophthalmol. Vis. Sci. 2018;59(9):1223.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Oxygen-Induced Retinopathy (OIR) is a widely used rodent model of ischemic retinopathy. The percentages of retina with vaso-obliteration (VO) and neovascularization (NV) areas are the two primary parameters in this model. Currently, these areas are calculated using manual segmentation, a tedious and time consuming procedure. We have developed an automated pipeline and the purpose of our study was to extend the algorithm and create a fully deep learning, web-based workflow for VO and NV segmentation.
1000 retina flatmount images of mice that went through the OIR procedure were obtained. Manual segmentation of VO and NV area were performed to serve as the ground truth for training. Two neural networks were trained separately for VO and NV. Due to the lack of manual segmentation of the total retina, we used kmeans to segment the 1000 images and manually picked the best 357 to use as the ground truth. Modified U-net architecture was utilized. The Dice coefficient was used to measure segmentation accuracy in the test set after training. A separate set of 37 images were manually segmented by four human experts. Each grader was set to the ground truth and the other 3 graders and the deep learning output were compared to the ground truth by Dice coefficients for the VO, NV and total retina regions. The percent of VO and NV regions were calculated and a linear correlation coefficient was calculated for each pairwise comparison.
Our model achieved similar range of correlation coefficients (range 0.890 - 0.955) to expert inter-human correlation coefficients (range: 0.889 - 0.951) for the percent area of VO, and achieved a higher range of correlation coefficients (range: 0.940 - 0.988) compared to inter-expert correlation coefficients (range: 0.925 - 0.972) for the percent area of NV. Our new retina segmentation model was able to handle the cases that our previous kmeans model failed. In addition, we deployed our pipeline into an open access, user-friendly website.
We created a fully automated, fully deep-learning based model for OIR segmentation, and achieved similar performance as human experts.
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