Automatic Segmentation Of Retinal Vessels In Fundus Images

Accurate automatic segmentation of retinal blood vessels plays an important role in the auxiliary screening of various eye diseases.However,due to the low contrast between the end of the blood vessel branch and the background in the fundus image,and the changeable shape of the cup and disc,the high-precision automatic segmentation of retinal blood vessels still faces many challenges.Traditional fundus image segmentation is manually completed by experienced ophthalmologists,but this process is vulnerable to subjective and objective factors.Therefore,it is of great practical significance and clinical value to study an accurate automatic segmentation method for fundus images based on modern computer aided technology.In this thesis,the following research has been carried out:(1)In order to solve the problems of insufficient feature extraction ability,serious loss of feature information and low segmentation accuracy of existing algorithms for color fundus retinal vascular images,an aggregated multi-scale integrated context model is proposed to further improve the accuracy of vascular segmentation.First,the low-level details of feature mapping at different scales are fully combined with high-level semantic information through the full scale jump connection from the encoding path to the decoding path.Secondly,a dense mixed expansion convolution block is designed between the encoder and the decoder to achieve accurate restoration of vascular details by obtaining richer context information.In addition,a compressed excitation block with residual connection is introduced into the decoder,which can strengthen the effective channel and suppress redundant information by adaptively adjusting the weight of each scale feature.Finally,the model performance experiment and module ablation experiment are carried out on the open data set,and the comparison experiment between AMIC-Net and different algorithms is carried out.The experimental results show that AMIC-Net has a higher recognition rate for blood vessels and is superior to many existing excellent retinal blood vessel segmentation methods.(2)In order to solve the problem of blood vessel discontinuity in fundus image segmentation,further reduce the number of model parameters,speed up the model convergence,and improve the generalization ability of the model,a lightweight dual path cascade retinal blood vessel segmentation algorithm is proposed.Firstly,a dual path cascaded network based on U-shaped codec structure is designed.In the coding and decoding part,a structured discarded convolutional block is designed to alleviate the over fitting problem of complex models and improve the generalization ability of models.Secondly,the depth separable convolution technique is introduced into the structured convolution block,which greatly reduces the parameters of the model;In addition,in the middle layer of the network,a hollow space convolution pyramid branch with residual connections is constructed,which effectively improves the information flow and achieves the goal of aggregating multi-scale information and expanding the receptive field.Finally,ablation and contrast experiments were carried out on the public dataset,and the generalization performance of the model was verified.The model parameters and computational complexity were compared and analyzed.The experimental results show that the proposed LDPC-Net segmented blood vessels have better connectivity,while the parameters of the model are only 1.95 M,which further shows that the proposed LDPC-Net is lightweight and efficient.