Research On The Method Of Blood Vessel Extraction Of Fundus Image Based On Multi-feature Fusion

Retinal fundus image(also called retinal angiography image in medical terminology)is a projected medical image of the surface of human eyeball.Doctors can observe the retina and its internal features through such images,including the shape of vascular trees,the spatial location of blood vessels,the state of optic discs,etc.Retinal vessels can provide valuable information for the diagnosis of various diseases.Changes in vascular morphology and location can effectively reflect cataract,glaucoma,diabetes,vasodilation and other diseases.Therefore,the processing analysis and automatic recognition of retinal fundus images are of great significance for the diagnosis and adjuvant treatment of fundus vascular diseases.With the development of artificial intelligence,a large number of machine learning models and depth learning models have been applied to automatic diagnosis of medical images.Due to the lack of contrast between blood vessels and background,uneven illumination of background,and single features of training classification in some fundus images,the classification results obtained by the trained classification model are not precise enough,especially insensitive to the details of blood vessels in blurred fundus images with cataract and other diseases,and the accuracy of the extracted blood vessel tree and the edge accuracy need to be improved.In this paper,based on the combination of multi-feature fusion and depth learning model,the method of fundus image blood vessel extraction is studied.The main work is as follows:(1)Carry out image preprocessing on fundus color image data set provided by the hospital.The fundus color image is grayed out,enhanced and compared with channels to select images with clear features and complete pretreatment.(2)Through different experimental methods,various feature images are obtained: spatial shape segmentation features of blood vessels,edge features of blood vessels,texture features of blood vessels and gradient features of blood vessels.(3)Multi-space recombination of the gradient and texture features obtained in step 2 and the grayscale features obtained from the original image,and then use the recombined new RGB image to train the fully convolutional neural network(FCN)model to obtain the fundus blood vessel Split graph.Finally,morphological processing,that is,high-cap filter processing and thresholding processing is added to obtain the final blood vessel segmentation image.(4)Multi-feature fusion processing is carried out on the obtained three feature maps of vascular spatial shape feature,texture feature and edge feature by using PCA method to obtain a comprehensive feature map containing fundus vascular texture,edge and spatial shape,and then FC-Dense CRF model is trained by using the comprehensive feature map to obtain a segmentation map of fundus image blood vessels.Through experimental analysis and comparison,it can be seen that the two algorithms for recognizing and extracting blood vessels in this paper have achieved better results,and the accuracy index of the two models for fundus image blood vessel segmentation is improved compared with the existing algorithms.The blood vessel extraction effect for blurred fundus images with cataract is also improved.