Improving vessel connectivity in retinal vessel segmentation via adversarial learning

Despite having achieved human-level performance in retinal vessel segmentation, deep learning based methods still suffer from poor connectivity of vessels in the generated segmentation maps. Since most methods operate as pixelwise classifiers, the vessel structure is ignored during the optimization of the segmentation network. To address this problem, a novel framework is proposed to enhance the vessel connectivity by incorporating the vessel structure into the segmentation network. First, to obtain the structural priors, the vessel structural priors extraction module (VSPEM) is proposed; VSPEM employs the powerful feature extraction ability of the convolutional autoencoder. After being pretrained, the proposed VSPEM can be used to extract useful latent features from the ground truths, which perform as the structural priors in segmentation. Then, the segmentation network is enforced to generate results that follow the distribution of the learned priors via adversarial learning. We have validated our method on three publicly available datasets, i.e., the DRIVE, CHASE_DB1 and STARE, and the state-of-the-art experimental results achieved on the above datasets demonstrate the efficacy of the proposed framework. Moreover, we show that the proposed framework is independent of segmentation models and can further improve model performance on vessel connectivity without introducing extra memory or a computational burden.