Abstract
Retinal vessel segmentation based on deep learning requires a lot of manuallabeled data. That is time-consuming, laborious and professional. What isworse, the acquisition of abundant fundus images is difficult. These problemsare more serious due to the presence of abnormalities, varying size and shapeof the vessels, non-uniform illumination and anatomical changes. In this paper,we propose a data-efficient semi-supervised learning framework, whicheffectively combines the existing deep learning network with GAN andself-training ideas. In view of the difficulty of tuning hyper-parameters ofsemi-supervised learning, we propose a method for hyper-parameters selectionbased on particle swarm optimization algorithm. To the best of our knowledge,this work is the first demonstration that combines intelligent optimizationwith semi-supervised learning for achieving the best performance. Under thecollaboration of adversarial learning, self-training and PSO to select optimalhyper-parameters, we obtain the performance of retinal vessel segmentationapproximate to or even better than representative supervised learning usingonly one tenth of the labeled data from DRIVE.