Research On Fundus Blood Vessel Segmentation Method Based On Multi-Task Learning

In recent years,the development of medical image processing has entered a new era along with the improvement of computing power and the fast development of artificial intelligence.Accurate segmentation of fundus vessels of the OCT images can effectively assist professional physicians to make accurate judgments on patients’ ocular lesions,and it is an indispensable and important step in the diagnosis and treatment of ophthalmic diseases.However,the fundus images often present the characteristics of multiple noise,high heterogeneity,class unbalanced issue,etc.Therefore,accurate segmentation of blood vessels from fundus images is still a very challenging task.Aimed at the problem of insufficient precision of traditional fundus vessel segmentation methods,a new fundus vessel segmentation framework was proposed in this thesis.And the main research contents of this thesis are as follows:Firstly,analyzing the importance of the fundus vessel segmentation task and the characteristics of blood vessels in fundus images,an improved U-net segmentation network is proposed to improve the feature extraction capability of the segmentation network for fundus vessels and enhance the overall segmentation performance of the proposed framework,which is used to act as the base segmentation network of the proposed multi-task segmentation network.Experiments have been carried out on the public DRIVE data set.The recall and F-1 value indicators have got promotion improvement,and the response time does not increase significantly.The proposed improved U-net network could improve the segmentation accuracy of the fundus vessels with a lightweight network.Secondly,the conventional single-task learning method segments the thick and the thin vessels in fundus images respectively,resulting in the segmentation network neglects most of the rich relevant information between the two kinds of vessels,which could reduce the feature extraction performance of the network on the fundus vessel features and affects the segmentation accuracy of fundus vessels.Aimed at this issue,this thesis proposes a multi-task learning segmentation method to segment the two kinds of vessels respectively,and the experimental results show that the proposed segmentation method can enhance the detection and extraction ability of these two blood vessel features,and enhance the segmentation performance of the segmentation network.Thirdly,analyzing the different distributions of the thick and the thin vessels in fundus images,as well as the serious class imbalance issue in the pixel ratio of the two different types of vessels,the traditional cross entropy loss function will lead to the small gradient in the segmentation network of fine vessels,which cannot accurately extract the features of fine vessels in fundus images.In view of the above problems,two kinds of loss functions are designed in this thesis to adapt to two different segmentation tasks.They can be used in the training of multi-task learning network module,which can complete the full detection and accurate segmentation of the fine vessels in fundus images.Finally,in the segmentation results of multi-task learning segmentation module,there are segmentation crossover and misjudgment in the junction of thick and fine blood vessels and other similar locations.And the direct combination of the two kinds of vessels segmentation results will lead to inaccurate final segmentation results Therefore,a fusion network module is designed in this thesis to adjust the coarse segmentation results of the two kinds of vessels.Experimental results on DRIVE,STARE and CHASE＿DB1 data sets shows that the vascular recall and F-1 value are the highest compared with other recent segmentation methods.Moreover,excellent segmentation performance is also shown in terms of global accuracy and specificity,and the vessel structure in fundus images could be completely and accurately detected.