Research On Retinal Vessels Segmentation Method Based On Supervised Learning

The retinal blood vessels are continuous and dendritic structures,and their branches start from the optic disc,and the width of the blood vessels decreases as they move away from the optic disc.Meanwhile,the optic disc is the confluence of the main blood vessels,which is larger and has obvious contrast compared with the blood vessel area here,and the pathological information such as hard exudate in the pathological retina will have an impact.Therefore,the challenges of retinal vessels segmentation include low capillary and background contrast,mis-segmentation of optic disc boundaries,and interference with pathological spots.At this stage,a large number of techniques based on computer-assisted retinal vessels segmentation methods have emerged.The use of computer software to evaluate their geometric parameters can effectively separate information such as blood vessels,lesions,maculars,and optic discs,which can promote medical experts’ systematic and consistent retinal blood vessel images Therefore,this paper proposes three types of algorithms: a multi-scale filtered blood vessel segmentation algorithm based on traditional machine learning,an adaptive scale information ushaped network segmentation algorithm based on deep learning,and a generation adversarial network segmentation method that fuses contour information.Its main work is as follows:(1)Retinal vessels segmentation algorithm with multi-scale filtering.Considering that the existing traditional machine learning algorithm has a single scale to extract vessels features,the difference of vessels features between different data sets is large,and it is easy to exist the phenomenon of wrong segmentation of optic disc and focus,and unreasonable microvascular segmentation.This paper uses multi-scale morphological filtering,matching filtering,Frangifiltering and 2D-Gabor filtering to track the direction of blood vessels,and obtains feature vectors that reflect the essence of target classification as much as possible.Considering the small amount of retinal image data,AdaBoost classifiers with good robustness and generalization are selected for classification decision.Finally,post-processing algorithms such as connected domain information are used to remove noise and other irrelevant information.(2)U-shape retinal vessels segmentation algorithm with adaptive scale information.Aiming at the defects of existing retinal vessels segmentation models,such as model solidification and artificial design features,deformable convolution is added to the coding structure of U-net network to better capture the morphological structure and scale information of blood vessels.At the same time,considering that the deformable convolution will have the possibility of mis-division,the attention mechanism is introduced in the decoding structure,which enables the decoding structure to capture the deeper semantic information of the target more accurately.Because of the characteristics of Unet structure,the decoder can effectively combine shallow texture information to improve the segmentation accuracy,thus reducing the loss of overall texture information.(3)Generation adversarial network segmentation method that fuses contour information.Considering that most of the algorithms are not robust in microvascular segmentation and some of the blood vessels are easily classified as background information after binarization,this paper improves the condition to generate the loss function of the generation network,so that the network can learn more robust retinal vessels contour information.At the same time,the distance penalty is used to constrain the distance between the gold standard information and the generated image.In addition,this paper makes use of channel and different scales attentional mechanism in generator network.The experiment shows that the two attentional mechanisms and contour loss entropy can be used simultaneously to not only segment more microvascular information reasonably,but also suppress irrelevant information such as background noise.