Structure Semantic Awared Retinal Vessel Segmentation Method

Accurate retinal vessel segmentation is essential to assist doctors in screening early fundus diseases.With the help of computer automated retinal vessel segmentation technologies,segmentation results of retinal vessels can be quickly obtained,which reduces the time consumption for ophthalmologists to manually identify retinal vessels.These technologies can greatly assist and improve the diagnosis procedure and have a good application prospect.This thesis explores the design of multi-task convolutional neural network from the perspective of the representation and utilization of retinal vascular structure and proposes two frameworks for retinal vessel segmentation task and retinal artery-vein segmentation task respectively.The main work is as follows:(1)For the retinal vessel segmentation task,this thesis proposes a skeleton-guided retinal vessel segmentation framework which introduces vessel centerline extraction as an auxiliary task to help better describe the vascular topology structure.Meanwhile,an Self-Adaptive Feature Fusion(SAFF)module is designed to integrate the features of the centerline extraction network and the retinal vessel segmentation network to improve structural integrity of the segmented retinal vessels.(2)For the retinal artery-vein segmentation task,this thesis proposes an artery-vein segmentation framework based on the relation representation of vessel structure,which takes vessel segmentation as an auxiliary task.A vessel structure relation module is designed and embedded between the two tasks.The module views the features of auxiliary task as structure reference.The similarity between vessel pixels is measured based on their high dimensional features,which helps to better distinguish the vessel category and improve the semantic consistency.In training phase,the gradient density based weighted cross-entropy loss is adapted to multi-class mode to alleviate the problem of data imbalance of arteries and veins.The two frameworks are evaluated by ablation experiments on several public fundus image datasets respectively to verify the efficacy of each component and are compared with other advanced retinal vessel segmentation and artery-vein segmentation methods.Experiments show that the two frameworks proposed in this thesis have achieved leading effect,and have strong versatility and robustness for end-to-end segmentation networks.