Unsupervised Domain Adaptation with Self-selected Active Learning for Cross-domain OCT Image Segmentation

Abstract

Segmentation of optical coherence tomography (OCT) images of retinal tissue has become an important task for the diagnosis and management of eye diseases. Deep convolutional neural networks have shown great success in retinal image segmentation. However, a well-trained deep learning model on OCT images from one device often fail when it is deployed on images from a different device since these images have different data distributions. Unsupervised domain adaptation (UDA) can solve the above problem by aligning the data distribution between labeled data (source domain) and unlabeled data (target domain). In this paper, we propose an UDA adversarial learning framework with self-selected active learning. The framework consists of two parts: domain adaptation module (DAM) and self-selected active learning module (SALM). The DAM learns domain-invariant features (i.e., common features) gradually to narrow the distribution discrepancy between two domains. The SALM introduces the target data into source domain through discrepancy method and similarity method, which promotes the DAM to learn unique features of target domain. Extensive experiments show the effectiveness of our method. Compared with the state-of-the-art UDA methods, our method has achieved better performance on two medical cross-domain datasets.