Retinal Vessel Segmentation Method Based On Automatic Labeling And Multi-task Learning

The characterization of retinal vessels is closely related to the early diagnosis of many comprehensive diseases,and the analysis of vessel structure is an important basis for disease diagnosis and treatment.Computer-based automatic retinal vessel segmentation technology can reduce the resource consumption and save the labor costs of professional doctors.It is a hot topic in the field of computer vision and medical image analysis.Based on the above background,this paper studies three aspects: label extraction,vessel topology enhancement and multitasking learning of deep learning retinal vessel segmentation methods:（1）Automatic vessel label extraction algorithm for retinal images.Recently,supervised deep learning methods have achieved great success in the field of retinal vessel segmentation,but the complexity of retinal images has led to the high cost of manual annotation.Therefore,annotation cost of large amount of data has been a significant obstacle to real applications of these approaches in the field of retinal vessel segmentation.In this paper,we first construct the vessel structure features detected by the optimally oriented flux filter.Then the idea of stratification is used to obtain the thick vessels and thin vessels of the retinal image through adaptive threshold method and skeleton tracking respectively.After modification and fusion of thick vessels and thin vessels,the label with thick and thin vessel classification is finally obtained to support the training of the deep learning model.（2）Adaptive topology enhanced deep learning segmentation method based on automatic labeling.In order to solve the common problems of vessel topological continuity and thin vessel loss in the segmentation results of existing methods,an adaptive topology-enhanced loss function is proposed based on automatic vessel label.The thick and thin vessels are enhanced through differentiation,which makes up for the disadvantage that the binary cross entropy loss function only considers a single pixel and ignores local information.The experimental results on DRIVE,STARE,CHASE＿DB1,HRF and IOSTAR datasets showed that the adaptive topology enhanced loss function improves the continuity of vessel segmentation results and reduces the leakage rate of thin vessels.（3）Multi-task deep learning method.Most of the errors in the existing deep learning retinal vessel segmentation methods occur at the edge of vessels.In order to solve this problem,this paper introduces a multi task learning mechanism to learn the vessel edge detection task and vessel segmentation task in parallel,so as to make use of the complementarity between them to jointly optimize the vessel segmentation results.To integrate the two complementary tasks,this paper proposes a fusion regular module to fuse the characteristics of the two branches of the network and regularize the parameters in the network.In addition,a multi-supervision module is proposed to introduce the multi-scale idea into the supervision training of the model,and the supervision is performed after each decoder output of the model,which enhances the robustness of the network.Experiments show that this method has certain advantages in segmentation accuracy,and the results of vessel segmentation with clear boundaries can be observed.