Automated Deep Learning Quantification of Avascular Area and Intravitreal Neovascularization in Retinal Flatmounts of Rodent Oxygen-Induced Retinopathy Models

Purpose: To develop a single deep learning model that quantifies the retinal avascular area (AVA) and intravitreal neovascularization (IVNV) in rodent oxygen-induced retinopathy (OIR) models. Methods: A U-Net-based model was developed to analyze AVA and IVNV in lectin-stained retinal flatmounts. The model was trained on 325 images (267 mouse and 58 rat) and evaluated on an independent test set of 37 images (18 mouse and 19 rat) annotated by human graders. We assessed intergrader reliability and agreement at metric and pixel levels. Mouse pixel-level performance was also compared with a previously published model. Results: Intergrader reliability was high for percent AVA (mouse intraclass correlation coefficient [ICC] = 0.840; rat ICC = 0.971), moderate for rat percent IVNV (ICC = 0.509), and low for mouse percent IVNV (ICC = -0.082). Metric-level correlation was strong in rat OIR (percent AVA r = 0.979; percent IVNV r = 0.943) and for mouse percent AVA (r = 0.957), but weak for mouse percent IVNV (r = 0.265). The Dice similarity coefficient was high for total retina (TR)/AVA and moderate for IVNV (rat: TR = 0.983, AVA = 0.924, IVNV = 0.612; mouse: TR = 0.975, AVA = 0.912, IVNV = 0.601). In mouse OIR, the Dice similarity coefficient matched or exceeded the previously published model (AVA = 0.912 vs. 0.887; IVNV = 0.601 vs. 0.559). Reviewers selected the IVNV mask created by the model in 83.3% of qualitative comparisons. Conclusions: Our deep learning model supports automated rat OIR analysis while maintaining mouse performance and may improve reproducibility of OIR measurements. Translational Relevance: Rodent OIR models are necessary to understand retinopathy of prematurity (ROP) pathophysiology. Our deep learning model effectively quantifies features of ROP recapitulated by both mouse and rat OIR.