Fully automated, deep learning segmentation of oxygen-induced retinopathy images

• DOI: 10.1172/jci.insight.97585
• Publication Date: 2017-12-20T16:01:44Z
• Authors: 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
• Keywords: Male; Observer Variation; Microscopy, Confocal; 610; Mice, Transgenic; Retinal Neovascularization; Mice, Inbred C57BL; Oxygen; Disease Models, Animal; 03 medical and health sciences; Deep Learning; 0302 clinical medicine; Image Processing, Computer-Assisted; Animals; Female; Algorithms

Oxygen-induced retinopathy (OIR) is a widely used model to study ischemia-driven neovascularization (NV) in the retina and serve proof-of-concept studies evaluating antiangiogenic drugs for ocular, as well nonocular, diseases. The primary parameters that are analyzed this mouse include percentage of with vaso-obliteration (VO) NV areas. However, quantification these two key variables comes great challenge due requirement human experts read images. Human readers costly, time-consuming, subject bias. Using recent advances machine learning computer vision, we trained deep neural networks using over thousand segmentations fully automate segmentation OIR While determining area VO, our algorithm achieved similar range correlation coefficients expert inter-human coefficients. In addition, higher compared inter-expert neovascular tufts. summary, have created an open-source, automated pipeline values images networks.