Research On Retinal Blood Vessel Segmentation Based On Deep Learning

The fundus retinal image is the only vascular image that can be observed by non-invasive means and has high research value in clinical practice.Ophthalmology and internal medicine experts can accurately judge a variety of diseases according to the different characteristics of retinal vessels.However,due to the influence of imaging equipment and illumination,there are many problems in the actual clinical application.Therefore,it is necessary to develop a high-precision automatic retinal vessel segmentation algorithm for clinical auxiliary diagnosis.In this thesis,based on the deep learning technology,the retinal vascular segmentation algorithm is studied.Based on the current research progress,this thesis proposes to use different improvement measures to improve the existing model.The main work of this thesis as follows:(1)Because the retinal blood vessel image has more details,it is easy to segment the retinal blood vessel at the end of the blood vessel by using the traditional neural network.Based on the U-Net network model as the backbone network,this thesis proposes an ASR-UNet network model of retinal vessels based on attention mechanism.The contrast experiment and ablation experiment show that the attention mechanism can greatly improve the performance of retinal vascular segmentation network and the segmentation accuracy of thiny vessels.The experimental results show that the ASR-UNet can on the DRIVE data set and CHASE＿DB1 data set respectively 0.9696 and 0.9637 accuracy.(2)Due to the limitations of the traditional U-shaped network,such as the limited information flow of the image in the model,poor anti-interference ability and so on.In combination with the improvement measures of U-Net model in the previous chapter,according to the characteristics of the model,Ladder Net and channel attention mechanism are combined,and multi-scale feature fusion module is used to fuse the information flow in multiple paths,and features of retinal vessels are extracted on multiple scales.The results of the comparative experiment and the ablation experiment show that all the improved measures in this paper can effectively improve the segmentation effect of the model,and the accuracy rates of 0.9698 and0.9674 are achieved in the DRIVE data set and STARE data set,respectively.In the DRIVE dataset,CHASE＿DB1 dataset and STARE dataset show that the neural network model proposed in this thesis can segment retinal vessels effectively,and compared with the retinal vessels segmentation network in recent years,the neural network model proposed in this thesis performs better in various indicators.At the same time,the improvement measures of the algorithm in the segmentation model provide ideas for the follow-up research.