Research On Retinal Vessel Segmentation In Fundus Image Based On The Bridge-Net

Some retinal diseases are of the most significant public concern worldwide due to the risk of irreversible blindness in recent years.Not only affecting the physiological functions of the human body,but these retinal diseases can also affect the stability and regularity of the morphology of the retinal vessels.Therefore,the analysis of the morphology of the retinal vessels in the fundus image can assist in the diagnosis of some retinal diseases,and segment retinal vessels accurately and effectively are the premise of the description and analysis of vascular structures.Manual retinal vessel segmentation is tedious and time-consuming which encourages researchers to study computer-aided methods for automatic segmentation.Among these approaches,the application of patch-based deep learning methods is increasingly widespread.However,most of the patch-based methods only use a coincident patch to describe the target region,and ignore the influence of the context information of the target region.In order to overcome the shortcoming,we adopt the deep learning method to carry out the retinal vessel segmentation task.The main responsibilities of our method include:(1)We propose a novel deep network architecture named bridge-net to use the context information of the target region effectively.For each target region,bridge-net extracts two concentric patches with inclusion relationships to represent two types of descriptions of the target region with and without its context information.These two patches form an input sequence that fed into a convolutional neural network(CNN)structure to generate features with and without context information correspondingly.Then the bridge-net uses a recurrent neural network(RNN)to deliver the context information of the target region by fusing the features obtained by the CNN to make the blood vessel segmentation in the target region more accurate.(2)Considering the difficulty in segmenting blood vessels in different patches might be various,we design a patch classification algorithm to classify the extracted patches and propose a patch-based loss weight map to correct the imbalance between blood vessels and background based on the classification which makes the allocation of weights more targeted.(3)We evaluate our method on STARE,DRIVE,and CHASE_DB1 which has significant differences in vascular morphology.The results show that the bridge-net can use the context information effectively,and make the segmentation more accurate,and the patch-based loss weight map can further improve the segmentation result.At the same time,the variability in the segmentation results of these datasets obtained by the bridge-net is tolerable which means the segmentation result of the network is less affected by the dataset,and the segmentation is stable.Compared with the existing segmentation method,the segmentation result of bridge-net is better than most of the state-of-the-arts,and in the methods with stable effect,the bridge-net achieves the best result in AUC.