Deep learning for automatic diagnosis of fundus diseases on optical coherence tomography images

Abstract

Abstract Background There has been increasing attention on the use of deep learning systems and transfer learning to identify disease features and progression. In this study, we aimed to apply deep learning algorithms to Optical coherence tomography (OCT) images to quickly identify common referable fundus diseases. Method A total of 24000 OCT images (test 1) retrospectively acquired from the Kaggle database with age-related macular degeneration, choroidal neovascularization, central serous chorioretinopathy, diabetic macular edema, diabetic retinopathy, drusen, macular hole, and NOR were used to develop the model. Images were split into training, validation, and testing sets. The convolutional neural networks ResNet101 and DenseNet121 were trained to classify images. The area under the receiver operating characteristic curve (AUC), accuracy, and F1 score were calculated to evaluate the performance of the models. A total of 800 OCT images (test 2) diagnosed with the above eight diseases were collected from the Zhengzhou Eye Hospital to retest the accuracy of the models. Results ResNet101 performed better than DenseNet121 did. The classification performance in terms of accuracy and F1 score of ResNet101 were 0.9398 and 0.9360, respectively, in test 2. The AUC of ResNet101 for the eight diseases based on test 2 were 0.9956 (macro-average) and 0.9932 (micro-average). When using DenseNet121 in test 2, the accuracy was 0.7130, and the F1 score was 0.7116. The macro-average AUC was 0.8519, and the micro-average AUC was 0.8366. Conclusions Convolutional neural network ResNet101 and transfer learning showed good performance in discriminating between OCT images. As a promising adjunctive tool, our model can provide rapid provisional diagnosis for patients with common referable fundus diseases.