Research On Fundus Vascular Segmentation Based On Neural Network

Abnormal changes in fundus vessels reveal the severity of many non-ophthalmic diseases such as diabetes,hypertension,arteriosclerosis,cardiovascular disease and stroke.Therefore,vascular segmentation is an important prerequisite for the diagnosis of many diseases.Usually the extraction of blood vessels is done manually by a doctor,but this is a time consuming task and relies heavily on the professionalism of the doctor.With the rapid development of image processing technology in the medical field,automatic segmentation and detection of fundus vessels has become a possibility.However,due to the complex features of the human fundus structure,the number of capillaries is large,and the segmentation is difficult,and the current segmentation technique still needs to be improved in the segmentation accuracy.On the basis of reading a large number of literatures,this thesis proposes an improved U-Net convolutional neural network to segment the fundus image.The fundus image of the public database DRIVE is used for training and detection,and finally the better vascular tree is segmented.The segmentation method is divided into two parts: preprocessing and blood vessel segmentation.In this thesis,the green channel image of RGB image is selected,and the fundus image is subjected to CLAHE processing and gamma correction to improve the contrast between the blood vessel tree and the background.Then use a two-dimensional Gabor filter to filter,filter out most of the background noise,and enhance the fundus blood vessels,providing a high-contrast picture for subsequent blood vessel segmentation.This thesis introduces the structure and principle of neural network,convolutional neural network,FCN and U-Net network in detail,and builds an improved U-Net convolutional neural network structure,which includes eight-layer neural network.The second to seventh layers are hidden layers and contain two downsampling layers.The network is trained and tested using the fundus image of the DRIVE database,and the resulting image can completely segment the vessel tree.Compared with the methods of the predecessors,the three vascular segmentation indexes of this thesis are better than the previous literature,in which the accuracy is 0.9691,the sensitivity is 0.8522,and the specific is 0.9821.At the same time,the method of this thesis has the following three advantages: 1)The blood vessel edge is smooth and without defects,the blood vessel thickness is similar to the original image,and the blood vessel fracture is very little.The whole blood vessel tree is very similar to the blood vessel tree segmented by medical experts.2)The capillaries can be well identified and segmented,and the segmented capillaries are smooth and complete with less breakage.3)The blood vessels of the optic disc,the optic cup,and the macula can be well segmented,and these fundus structures are not mistakenly recognized as blood vessels.