Deep Learning-based Color Fundus Retinal Model Blood Vessel Segmentation And Non-deterministic Research

As a disease reactor that we can observe directly,retinal vessels can help doctors diagnose various diseases.Due to the development of science and technology,it is very convenient to obtain fundus images through medical technology.The damage of fundus blood vessels can be clearly detected by instruments.However,due to the complexity between blood vessels,the contrast of capillary area is low due to the influence of strong illumination background,and the accuracy of computer-aided segmentation of retinal blood vessels is low.Recently,deep learning has been widely used in the field of medical images.A large number of retinal vascular segmentation algorithms and technologies emerge in endlessly.At the same time,it promotes ophthalmologists to comprehensively evaluate the quality of fundus images.Therefore,this paper proposes three kinds of algorithms:vascular segmentation algorithm based on edge enhancement and feature de dryness,multi-scale fusion network segmentation algorithm based on attention perception,and uncertain research of vascular segmentation network based on Bayesian deep learning.The main work is as follows:·Vessel segmentation algorithm based on edge enhancement and feature de dryness.Because the common CNN model is easy to lose the high-frequency information of the image,it is easy to produce the phenomenon of segmentation error.Therefore,this paper selects the combination of traditional machine learning and CNN algorithm,and selects Canny edge extraction operator to obtain the edge prior information of fundus image.In addition,we also use feature denoising blocks to reduce the noise hidden in low-level features and reduce the impact on the performance of network segmentation.·Multiscale fusion network segmentation algorithm based on attention perception.The traditional convolution operation is not ideal for capillary segmentation.Therefore,this paper proposes a location and channel attention module to enable the model to effectively extract features in multi-dimensional space and further improve the segmentation performance.At the same time,this paper also proposes adaptive multi-scale dense blocks to extract the features of different scales of the image,so as to improve the segmentation results.And this paper also uses multi-scale fusion module to reduce the loss of semantic information and detail information.·Research on uncertainty of blood vessel segmentation network based on Bayesian deep learning.Based on the multi-scale fused blood vessel segmentation algorithm based on attention perception proposed in Chapter 4,the uncertainty of the network is analyzed.It is used to help doctors judge the confidence of the prediction results of the model.Qualitative,quantitative and ablation experiments were carried out on two public fundus data sets.Experimental results show that the proposed model has better robustness than other algorithms,and each component effectively improves the segmentation performance of the model...