Semi-supervised Adaptive Weighted Network for CNV Typing in OCT Images

Choroidal neovascularization (CNV) is a wet age-related macular degeneration (AMD), which will seriously affect the vision of patients. Accurate CNV typing in OCT images plays an important auxiliary role in disease treatment. However, large number of noises may be introduced due to various factors such as imperfections of imaging devices, which affects the diagnosis of the disease. In addition, class imbalance will arise, result in decrease of generalization performance of the model. Considering that there are limited labeled data and large number of unlabeled data in clinical, this paper proposed the semi-supervised adaptive weighted network (SSAW-Net) for accurate typing of CNV. Firstly, Saliency-Noising data augmentation method is proposed to generate images which the noise is introduced in the salient regions, to force the model to learn robust features to the noise. In order to make full use of the unlabeled data, we adopted a semi-supervised network architecture based on Mean-Teacher framework as the backbone. In the proposed network, the adaptive weighted strategy based on meta learning is developed to assign the weights to the samples adaptively during the training process. This method can avoid the class imbalance problem. We conducted extensive experiments on private and public datasets to demonstrate the effectiveness of our method.