Research On Vascular Image Segmentation Technology Based On Deep Learnin

The initial diagnosis of retinal diseases relies heavily on the analysis of fundus images by ophthalmologists.Diseases similar to diabetic retinopathy,macular degeneration,and hypertension all present different features in the vascular morphology of the retina.However,due to the complex structure of the retina and the different degrees of lesions in various diseases,a large amount of human and material resources are required to perform large-scale screening and manual diagnosis,and therefore,it is easy to miss and misdiagnose.Currently,computer technology to assist in the diagnosis of retinal diseases is one of the current focuses of medical research and computer direction.In this paper,we focus on algorithms related to retinal image vascular segmentation,as follows:(1)Study the imaging principle of retinal images and image pre-processingThrough the study of retinal imaging principle and fundus image acquisition,we deeply understand the characteristics of fundus images and lay a solid foundation for the subsequent image preprocessing work.Fundus images are characterized by high noise and low contrast,and in order to make the subsequent network training results more optimized,the original images need to be preprocessed.In this paper,we mainly use grayscale transform,normalization,CLAHE,gamma transform and data augmentation to preprocess the original fundus images.The data set used in this paper is the DRIVE public data set,which contains 20 training sets and 20 test sets.Due to the small amount of data,data augmentation is needed,and the images are proposed to be cropped by random slicing to avoid overfitting or gradient disappearance of the subsequent network due to data scarcity.(2)Retinal vessel image segmentation algorithm based on U-net and Bi-FPN fusionThe backbone network used in this paper is the U-net network,which is very suitable for semantic segmentation of complex medical images due to its two features of left-right symmetry and bridging contextual features with jump connections.The original U-net network uses jump connections to achieve effective transfer of features and avoid problems such as loss of feature detail information due to downsampling.The Bi-directional Feature Pyramid Network(Bi-FPN)is proposed in the Efficient-Det network,which is more effective than jump connections in transferring features.The purpose of introducing the bi-directional feature network is to improve the feature extraction efficiency of the backbone architecture and each level,and to enrich the feature vector so as to achieve the fusion of low-level fine-grained features and high-level semantic features.Experiments show that the improved U-net network has some improvement over other methods on open datasets.In this paper,using the improved U-net network to segment vessels,the segmentation results achieved 80.65% Sensitivity,97.68% Specificity,95.51% Accuracy,83.51% Precision,and 97.87% AUC at the pixel level.The results were obtained with good experimental results.The experimental results also show that the algorithm of this paper has better performance on the vascular segmentation task when compared with other algorithms on the same data set and with the same evaluation parameters.