Abstract
Purpose :
Diabetic retinopathy, a leading cause of adult blindness, is a complex and incurable ocular disease driven by vascular damage in the retinas of patients with diabetes. Although the disease usually affects both eyes of a patient, the pathophysiology of the bilateral versus unilateral presentation remains unclear. The present investigation seeks to understand if fundus images of one diseased eye contain features that can be meaningfully learned by deep neural networks to indicate bilateral or unilateral diabetic retinopathy at the patient level.
Methods :
The public EyePACS diabetic retinopathy dataset was queried and found to contain retinal images from 3,811 bilateral and 1,500 unilateral diabetic retinopathy patients. The eye with the greater severity was selected, ties broken randomly, yielding a dataset of 5,311 total images split representatively into 4,248 (80%) images for training and 1,063 (20%) images for testing. The first approach involved transfer learning of convolutional neural networks pretrained on ImageNet, including AlexNet, VGG-16, and ResNet-18. The second approach utilized a non-pretrained ResNet-18 with a modified two-channel output layer as the network architecture. Model performance was evaluated with accuracy, precision, recall, and F1 score.
Results :
Pretrained AlexNet, pretrained VGG-16, pretrained ResNet-18, and non-pretrained ResNet-18 were implemented on PyTorch and fitted with the one-cycle learning rate policy on the training set. The pretrained networks were fine-tuned for 10 epochs, while the non-pretrained network underwent training for 20 epochs. Evaluating the four models on the test set resulted in prediction accuracies of 73.38%, 70.46%, 72.34%, and 71.40%, respectively. While the pretrained AlexNet resulted in the highest F1 score at 0.8404, the pretrained ResNet-18 resulted in the highest precision at 0.7516. The evaluation metrics for all four models are provided in Table 1.
Conclusions :
To our knowledge, this is the first study to demonstrate prediction of unilateral or bilateral diabetic retinopathy based on a single fundus photograph. We conclude that artificial intelligence is capable of performing this classification task with reasonable accuracy, suggesting the existence of latent features within retinal images that may yield insights into how diabetic retinopathy manifests globally in the patient.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.