Application of deep learning to quantify vascular tortuosity in mouse models of oxygen-induced retinopathy

Abstract

Purpose : Increased retinal vascular tortuosity has been quantified in retinal images of humans with retinopathy of prematurity (ROP). The oxygen-induced retinopathy (OIR) model of ischemic retinopathy mimics hallmark features of ROP, including initial ischemia followed by neovascularization. The purpose of this proof-of-concept study was to develop a semi-automatic deep learning algorithm that can be used to investigate whether there is a correlation between retinal vascular tortuosity and disease activity in OIR mice, analogous to what is observed in infants with ROP. Application of this algorithm to images of OIR retinas aimed to characterize vascular tortuosity as a novel outcome measurement in the OIR model.

Methods : OIR was induced in C57BL/6J mice via hyperbaric oxygen exposure from postnatal day 7 (P7) to P12. Retinal flat-mounts of P17 OIR mice were manually segmented for superficial vessels by 4 graders and validated using a subset of images. The optic disc of each image was manually demarcated prior to algorithm input. Using a previously pre-trained DL algorithm for calculating tortuosity index (TI) in retinopathy of prematurity (iROP-Assist), each segmentation and its corresponding disc center was used to generate a TI. Statistical significance between groups was defined as p ≤ 0.05 and determined using a Student’s T-Test.

Results : 50 flat-mount images representing normoxic (NOX) mice and 50 flat-mount images representing untreated P17 OIR mice were included in this analysis. The median tortuosity index for NOX and OIR images was 1.01 [IQR 0.0] and 1.04 [IQR 0.02] respectively (Figure 1), which represented a statistically significant difference (p < 0.01). Examples of paired real and segmented NOX and OIR images are shown in Figure 2.

Conclusions : The tortuosity index of retinal vasculature can be semi-automatically calculated for retinal images of OIR mice using a deep learning algorithm. Application of this algorithm to compare the TI for NOX and OIR mice at P17 demonstrated significantly greater tortuosity in OIR mice. Future studies may employ this tool to rapidly assess the effects of therapeutic interventions on the TI of the OIR phenotype.

This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.

 

Figure 1. Distribution of tortuosity indices (TI) and labeled medians for normoxic (NOX) vs oxygen-induced retinopathy (OIR) images

 

Figure 2. Real and segmented pairs of NOX and OIR images and their corresponding tortuosity indices