A Novel Consistency Regularization Scheme for Retinal Lesions Semi-Supervised Segmentation

Diabetic retinopathy (DR) is a significant complication of diabetes, which is the leading cause of loss of vision. The diagnosis of DR usually depends on red lesions (microaneurysms and haemorrhages) and white lesions (soft and hard exudates) in fundus images. Thus, the automatic and accurate segmentation of retinal lesions is an essential work for diagnosing DR. Deep convolutional neural networks (CNNs) have proven their superiority in many lesion segmentation tasks. However, current DR segmentation models require a large amount of dense pixel-level annotation to train, which is expensive and time-consuming. In this paper, we proposed a novel consistency regularization semi-supervised network (CR-Net) for retinal lesions segmentation. For labeled examples, we use a super vised mode to train the shared en coder and main de coder. For un labeled examples, we de sign a new consistency regularization loss, which fuses pixel - level and structure - level consistency to rectify the uncertainty pixels during training. We evaluate the CR-Net on the publicly retinal lesions segmentation dataset. The metric of mIoU, F-score, and G-mean score are 0.7019, 0.5930, and 0.7212 on the IDRiD dataset with the labeled data rate of 50%. An ablation study was also conducted to analyze the effectiveness of each module. The experiment results prove that the CR-Net achieves better retinal lesions segmentation performance than other semi-supervised methods.