Research On Retinal Vessels Segmentation For Fundus Images Based On Fully Convolutional Networks

Retinal blood vessels are the only blood vessels in the human body that can be observed non-invasively,which is of great value and significance for clinical medical diagnosis.Early symptoms of cardiovascular diseases such as diabetes and high blood pressure can change the diameter,direction and spatial distribution of retinal blood vessels,and even neovascularization.However,due to the limitations of the fundus camera device,the poor quality of fundus image,where exists uneven illumination,low contrast of the blood vessel tip and background,and the complexity of space distribution of the vessels as well as the interference of lesions makes manual segmentation very time-consuming and laborious,subjected to subjective factors.Therefore,it is very important to use computer-aided diagnosis technology to automatically extract retinal vessels.Experts and scholars at home and abroad have studied various methods to achieve automatic segmentation of retinal vessels.With the development of deep learning techniques,convolutional neural networks are also used to achieve retinal vessel segmentation.In this paper,two retinal vessel segmentation algorithms based on fully convolutional networks are designed to segment retinal vessels from fundus images:1)A patch-wise retinal vessel segmentation algorithm is designed in this paper.Firstly,the fundus image is converted to gray scale and normalized.Then the Contrast Limited Adaptive Histogram Equalization method is used to improve the contrast between blood vessels and background,and the image is cropped into patches by sliding window technology.Partially overlapped small blocks are input into the convolutional neural network.In this paper,a fully convolutional network based on encoder-decoder structure is designed.Residual dense block is used,which can fully extract image features.And ordinary convolution is replaced by dilated convolution in the networks,which increases the receptive field of networks without increasing the number of parameters.The image patches are passed through the trained network and spliced into a complete score map and then the final segmentation result is obtained by the Otsu threshold segmentation method.Experimental results demonstrate that the patch-wise method has certain robustness in the case of central vessel reflex and lesion area interference,and it has higher accuracy and sensitivity than other existing methods,which can extract more vessels.2)An end-to-end retinal vessel segmentation algorithm is implemented.Each channel of the original image is normalized,and then the whole image is input into the convolutional neural network to obtain the segmentation result.In this paper,a U-Netlike encoder-decoder structure FCN model is designed with the residual dense block and the dilated convolution.In the training process,the model is pre-trained with the cropped image patches firstly.Then the model is fine-tuned with full size images.The experimental results show that the end-to-end method has better segmentation performance on the DRIVE and CHASE＿DB1 datasets,and is less time-consuming than the patch-wise algorithm.