Semi-supervised Learning Framework in Segmentation of Retinal Blood Vessel Based on U-Net

Medical image analysis of retinal blood vessels is of great value for the disease warning of diabetic retinopathy and the diagnosis of cardiovascular and cerebrovascular diseases. An accurate segmentation of the vascular tree in fundus images is essential for medical analysis of retinal images. In this paper, a semi-supervised framework for segmentation of blood vessels based on U-Net is proposed, aiming to abate the workload of data annotation. The framework includes three steps. Firstly, U-Net is trained with the enhanced ground truth labels; secondly, use the trained network to predict unlabeled data, and take the filtered prediction results as pseudo-labels; thirdly, combine data amplification and dropout strategies to update the training set. Repeating these steps until the predetermined iteration times is reached. Then we use the trained model to segment the retinal vessel images, and report the segmentation performance on the test set. Through comparison experiments with fully-supervised learning method, we find that the proposed framework has better performance than fully-supervised learning under the same amount of labeled data, which can improve the effect of blood vessel segmentation and reduce the workload of data labeling.