Retinal Vessel Segmentation Based On Deep Learning Algorithms

In recent years,with the increasing prevalence of ophthalmology diseases such as diabetic retinopathy and glaucoma,the quality of human life has been affected a lot.To diagnose ophthalmology diseases,doctors need to examine the retina area.There are many multi-scale optic nerves and retinal vessels on the retina.Retinal vessel segmentation is a significant and fundamental step to construct such a computer-aided diagnosis system,which can be used for early diagnosis of ophthalmology diseases.However,the existing segmentation methods remains challenging to segment the retinal vessels,especially in the capillary areas.In conclusion,the difficulties of retinal vessel segmentation can be summarized as three parts:(1)limited densely annotated data;(2)scale imbalance of retinal vessels;(3)lack of spatial structures prediction.Therefore,addressing these challenges,our works are the following:(1)Methods based on ‘divide and conquer’ strategy.We transfer the single retinal vessel segmentation problem into fine-grained objects segmentation problems including thick vessels,middle vessels and thin vessels.In this model,we first classify the vessel categories and then segment the same-sized vessels.Hence,it can relieve the impact of scale imbalance;(2)Methods based on ‘cascaded network’ strategy.Compared to the conventional postprocessing module,we employ the followed network to refine the retina vessels and propose a network followed network model.Thereby,the model including one input and two outputs can be trained in an end-to-end manner.Also,we can use auxiliary loss to train our model effectively,which can balance the two identical networks;(3)Methods based on ‘multi-path supervision’ strategy.To make full use of limited training data and extract enough multi-scale features,we design four supervision paths including one original path,two multi-scale paths and one rich feature path.Furthermore,based on the parallel paths and residual structures,we construct a novel convolutional module to achieve a more powerful feature representation.As a result,we are able to generate good segmentation results.We evaluated the proposed methods on three public databases including DRIVE,STARE and CHASE databases.The results showed the powerful ability of our models to segment vessels and the ablation studies demonstrated each motivation can achieve impressive performance gains.We also obtained the preliminary results in the section of future work to show the generalization and expanding capacity of the proposed models.