Applications of Deep Learning: Automated Assessment of Vascular Tortuosity in Mouse Models of Oxygen-Induced Retinopathy

ObjectiveThe purpose of our study was to develop a generative adversarial network (GAN) segment major blood vessels from retinal flat-mount images oxygen-induced retinopathy (OIR) and demonstrate the utility these GAN-generated vessel segmentations in quantifying vascular tortuosity.DesignDevelopment validation GAN.SubjectsThree datasets containing 1084, 50, 20 mice retina with various stains used ages at sacrifice acquired previously published manuscripts.MethodsFour graders manually segmented retinas OIR mice. Pix2Pix, high-resolution GAN, trained on 984 pairs raw manual segmentations, then tested 100 50 image held-out external test set, respectively. were as an input into algorithm (iROP-Assist) generate cumulative tortuosity index [CTI] for mouse eyes treated Aflibercept versus IgG control.Main Outcome MeasuresMean Dice coefficients compare segmentation accuracy between annotated maps. For control, mean CTIs also calculated both Statistical significance evaluated Wilcoxon Sign-Rank Tests (p≤0.05 thresholded significance).ResultsThe coefficient 0.75±0.27 0.77±0.17 set set. In proof-of-concept experiment, CTI generated 1.12±0.07 1.03±0.02 (p=0.003) 1.06±0.04 1.01±0.01 (p<0.001), respectively, demonstrating that rescued by when quantified segmentations.ConclusionsGANs can be accurately map flat mount images. These maps may evaluate novel metrics OIR, such CTI, have potential accelerate research treatments ischemic retinopathies.