Deep Learning for Automated Detection of Neovascular Leakage and Vascular Nonperfusion in Diabetic Retinopathy Using Ultra-widefield Fluorescein Angiography

Abstract

Purpose : Neovascular leakage and vascular non-perfusion are angiographic features of diabetic retinopathy. The goal of this study was to determine whether these features could be accurately classified by a deep learning algorithm.

Methods : We used retrospective fluorescein angiogram images captured on the Optos ultra-widefield platform between January 2009 and May 2018. Images were randomly split into training (60%), validation (20%), and test (20%) sets. A ResNet-101 convolutional neural network that was pre-trained on ImageNet was trained on fluorescein angiogram images. The neural network was configured to output predictions on the presence of neovascular leakage and vascular non-perfusion. Predictions were compared to the ground truth labeled by masked, trained graders.

Results : 1452 images from 342 subjects were included. 60% of images were from female subjects, and 79% of images were from patients with type 2 diabetes mellitus. Mean hemoglobin A1c was 8.1% (SD=2.1%). Mean visual acuity was logMAR 0.35 (SD=0.30). Neovascular leakage was present in 21% of images, and non-perfusion greater than a pre-determined threshold (77.5 mm²) predictive of development of proliferative diabetic retinopathy was present in 42% of images. The mean total non-perfusion area was 80.8 mm² (SD=63.3). The algorithm achieved an area under the curve (AUC) of 0.976 for neovascular leakage, and an AUC of 0.949 for detecting vascular non-perfusion. At operating points selected for high sensitivity, the algorithm achieved 97% sensitivity and 89% specificity for neovascular leakage, and 94% sensitivity and 79% specificity for non-perfusion. At operating points selected for high specificity, the algorithm achieved 75% sensitivity and 94% specificity for neovascular leakage, and 86% sensitivity and 86% specificity for non-perfusion.

Conclusions : A convolutional neural network was trained to recognize the presence of neovascular leakage and non-perfusion in ultra-widefield fluorescein angiography. Further research may help improve algorithm performance and better characterize the relationship between non-perfusion, neovascularization, and vision-threatening complications of diabetic retinopathy.

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