Semi-Supervised Automatic Segmentation of Layer and Fluid Region in Retinal Optical Coherence Tomography Images Using Adversarial Learning

Abstract

Optical coherence tomography (OCT) is a primary imaging technique for ophthalmic diagnosis due to its advantages in high resolution and non-invasiveness. Diabetes is a chronic disease, which could cause retinal layer deformation and fluid accumulation. It might increase the risk of blindness, and thus, it is important to monitor the morphology change of the retinal layer and fluid accumulation for diabetes patients. Due to the existence of deformation and fluid accumulation, the retinal layer and fluid region segmentation in the OCT image is a challenging task. Machine learning-based segmentation methods have been proposed, but they depend on a significant number of pixel-level annotated data, which is often unavailable. In this paper, we proposed a new semi-supervised fully convolutional deep learning method for segmenting retinal layers and fluid regions in retinal OCT B-scans. The proposed semi-supervised method leverages the unlabeled data through an adversarial learning strategy. The segmentation method includes a segmentation network and a discriminator network, and both the networks are with U-Net alike fully convolutional architecture. The objective function of the segmentation network is a joint loss function, including multi-class cross entropy loss, dice overlap loss, adversarial loss, and semi-supervised loss. We show that the discriminator network and the use of unlabeled data can improve the performance of segmentation. The proposed method is investigated on the duke Diabetic Macular Edema dataset and the POne dataset, and the experiment results demonstrate that our method is more effective than the other state-of-the-art methods for layers and fluid segmentation in the OCT images.