A two-stage unsupervised domain adaptation method for OCT image segmentation

Abstract

Deep convolutional neural networks (CNN) have achieved great success in segmentation of retinal optical coherence tomography (OCT) images. However, images acquired by different devices or imaging protocols have relatively large differences in noise level, contrast and resolution. As a result, the performance of CNN tends to drop dramatically when tested on data with domain shifts. Unsupervised domain adaptation solves this problem by transferring knowledge from a domain with labels (source domain) to a domain without labels (target domain). Therefore, this paper proposes a two-stage domain adaptation algorithm for segmentation of retinal OCT images. First, after image-level domain shift reduction, the segmenter is trained with a supervised loss on the source domain, together with an adversarial loss given by the discriminator to minimize the domain gap. Then, the target domain data with satisfactory pseudo labels, measured by entropy, is used to fine-tune the segmenter, which further improves the generalization ability of model. Comprehensive experimental results of cross-domain choroid and retinoschisis segmentation demonstrate the effectiveness of this method. With domain adaptation, the Intersection over Union (IoU) is improved by 8.34% and 3.54% for the two tasks respectively.