Abstract
Purpose :
The standard of care for diabetic patients is to have an eye exam or retinal imaging to assess for diabetic retinopathy (DR). In recent years there has been increased interest in developing artificial intelligence (AI) tools for optimizing screening of ocular imaging. In this retrospective study, we developed an AI-based approach for assessing the gradeability of images in a DR screening program.
Methods :
Non-mydriatic retinal images were gathered from diabetic patients who underwent imaging during a primary care or endocrinology visit at Northwestern Medicine from September 1, 2017 to June 1, 2021. The Eidon SLO Confocal Scanner (Centervue Inc., Fremont, CA) was used. Images were interpreted by Northwestern ophthalmologists for gradeability, presence and severity of DR, and the presence of non-DR pathologies. Following image interpretation, the TensorFlow and Keras platforms were utilized to train convolutional neural networks (CNNs) to assess image gradeability. Images were divided into training, test, and validation sets. Inception V3 and Xception V1 network architectures were evaluated, with network weights preinitialized using a network trained on ImageNet, a set of 14 million images used for computer vision research. Networks built with different combinations of hyperparameters (learning rate (LR), dropout rate (DropR), and dense layer output dimensionality (DLOD)) were compared by validation set performance.
Results :
1550 non-mydriatic retinal images from 575 patients (55% female, median age 58) were analyzed in this study. Northwestern ophthalmologists deemed 23.6% (366/1550) of these images to be ungradable. Of gradable images, ophthalmologists found 20.4% (241/1184) had DR of varying degrees and 26.9% (319/1184) had non-DR ocular pathology.
Comparing CNNs, validation set performance was maximized with (LR = 0.001, DropR = 0.3, DLOD = 256) for the Inception V3 network and (LR = 0.0001, DropR = 0.3, DLOD = 256) for the Xception V1 network. On the test set the Inception V3 network exhibited 87.5% accuracy (AUC 0.935) and the Xception V1 network exhibited 86.41% accuracy (AUC 0.929).
Conclusions :
It is feasible to develop CNNs that can assess the gradeability of non-mydriatic retinal images with a high degree of accuracy. AI-based frameworks such as these may enable more efficient identification of ungradable images in DR screening programs.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.