Improving retinal OCT image classification accuracy using medical pre-training and sample replication methods

Eye disease is a global health issue involving a large population, and it is necessary to develop high -accuracy diagnostic methods. Manual diagnosis has some challenges, and deep learning for automatic diagnosis is becoming a research focus. However, there are two problems in current deep learning used for medical image tasks: deficient network pre -training on ImageNet dataset consisted of natural scene images, which significantly differ from medical images; a discrepancy between sub -network for training and full network for prediction introduced by dropout. In this study, a medical pre -training method and a sample replication method were proposed to address the problems, respectively. The former involves network pre -training on a large-scale medical image dataset called RadImageNet, followed by transfer learning on a retinal OCT (Optical Coherence Tomography) image dataset. The latter uses JS (Jenson's Shannon) divergence between the two predictions yielded with two identical samples via sample replication as an additional loss function to enforce network's outputs consistency. The two methods were applied on widely used ResNet50, DenseNet121 and InceptionV3 networks, and the effectiveness of using each method and both the two methods to improve the network's performance was demonstrated by experimental results. Compared to those of the corresponding base networks, most the metrics of the networks using the methods were improved for all four sample categories, and the accuracy improvements of 3.76 %, 6.52 % and 8.63 % were obtained for above the networks, respectively. All the networks achieved an accuracy of about 95 % even working with limited target dataset and training resources.