UASSDA-Net: Uncertainty-aware Semi-supervised Domain Adaptation network for choroidal neovascularization segmentation

CNV segmentation in OCT images plays an important role in the diagnosis and treatment of CNV. However, the presence of uncertain factors such as differences in equipment or imaging protocols results in irregular distribution in some CNV images. It's difficult for the existing CNV segmentation methods to deal with this problem because they ignore the learning of uncertain knowledge, result in the limitation of the segmentation performance improvement. In addition, the lack of labeled data results in poor generalization. To address the above two issues, this paper proposes an uncertainty-aware semi-supervised domain adaptation network for CNV segmentation. First, difficulty cutmix method is developed to construct an uncertain knowledge base that contains the CNV areas segmented incorrectly due to the uncertain factors. Afterwards, a semi-supervise uncertainty-aware domain adaptation method is proposed to fully learn uncertain knowledge from the uncertain knowledge base, and introduced the learned uncertain knowledge into the segmentation model. It can make the model more robust to the impact of the uncertain factors, to further improve the segmentation accuracy. The experimental results demonstrate the effectiveness of the proposed method.