Purchase this article with an account.
David Le, Minhaj Nur Alam, Jennifer I Lim, Robison Vernon Paul Chan, Xincheng Yao; Deep Machine Learning for OCTA Classification of Diabetic Retinopathy. Invest. Ophthalmol. Vis. Sci. 2020;61(7):4104.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
This study is to test the feasibility of using deep machine learning for OCTA classification of diabetic retinopathy (DR).
A deep learning convolutional neural network (CNN) architecture VGG16 was employed for this study, illustrated in Figure 1. Transfer learning was implemented from pre-trained weights, optimized from the ImageNet dataset, for OCTA classification of control, NoDR and NPDR. Additional procedures, including data augmentation, early stopping and cross-validation, were implemented to prevent overfitting. Our dataset comprised of OCTA images from 16 control subjects, 12 diabetes mellitus without DR (NoDR), and 37 NPDR (mild, moderate, and severe stages) patients. Data augmentation in the form of rotations, flips, and zooming were implemented to increase the dataset size to 3930 OCTA images. Each model was trained with early stopping and converges within 70 epochs. To evaluate each trained model, a 5-fold cross validation method was employed. Each fold followed an 80-20 train-test split procedure. To determine the optimal number of layers fine-tuned in transfer learning, a comparison of the misclassification error was performed on all fine-tuned layers of the CNN, summarized in Figure 2. Evaluation metrics for the best performing model included diagnostic accuracy, sensitivity, specificity, and AUC.
With the last nine-layers fine-tuned, the CNN model achieved the best performance for OCTA classification of DR. The evaluation metrics for the best performing model was 87.27% diagnostic accuracy, 83.76% sensitivity and 90.82% specificity. The AUC metrics for binary classification for control was 0.97, NoDR was 0.98, and NPDR was 0.97.
Transfer learning enables automated OCTA classification of DR, promising an effective method to foster rapid screening and easy staging of DR.
This is a 2020 ARVO Annual Meeting abstract.
Figure 1. (a) The CNN that was trained for classification of DR. The box contains the nine layers that were fine-tuned. (b) Comparison of misclassification error for the number of fine-tuned layers.
This PDF is available to Subscribers Only