Research On Retinal Vessel Segmentation Algorithm Based On Deep Learning

Funduscopy can detect retinal diseases and even systemic diseases,such as diabetes and hypertension.The key to fundus examinations is to extract the retinal vessel structures from fundus images to help doctors diagnose diseases as well as suggest treatment options.Due to the development of imaging technology and the increase of population,the fundus image data is becoming more and more abundant,and there is a need to automatically segment the fundus images to extract the required vessel image data.Many methods have been proposed in order to segment the vessel images,which can be divided into traditional segmentation methods and deep learning methods.Traditional segmentation methods mostly rely on image processing techniques and mathematical methods,which are simple to use,Better interpretability,but have lower segmentation accuracy and require high image quality to avoid being affected by noise.Deep learning methods produce more accurate segmented blood vessel images,but the deep learning models are difficult to deploy on low-computational power platforms or mobile devices to solve practical problems due to their large number of parameters and slow running speed.In 2015,Ronneberger et al.proposed U-Net,a concise and effective structure that became the backbone for many deep learning based retinal vessel segmentation models.In this paper,in order to solve these problems,we propose an improved model based on U-Net:(1)using U-Net as the backbone network,optimizing the structure,reducing the number of channels and the number of downsampling;(2)in order to improve the characterization ability of the model,using convolutional blocks with parallel structure to replace the convolutional blocks with single-way structure in the original U-Net,and repeating the convolutional block multiple times in the encoder,which improves the information extraction capability of the encoder,thus improves the model performance;(3)using structural reparameterization,change the parallel-structured convolutional block into a single-path convolutional block in the model inference stage.Reducing the model parameters and the improving inference speed by reparameterize the 3 × 3 conv,1 × 1conv and batch normalization operations into a new 3 × 3 conv.To evaluate the segmentation performance and inference speed of the proposed model,six deep learning-based retinal vessel segmentation models are selected and compared on three of the most widely used retinal vessel datasets.The experimental results show that in DRIVE the new model play the best performer on SEN,AUC,and F1 metrics among five metrics,with 0.8224/0.9871/0.8245 respectively,and comparable to the best performance level on other metrics;in STARE play the best performance on SEN and AUC,with0.8473/0.9923 respectively;in STARE play the best performance on ACC and AUC,with0.9731/0.9874 respectively.In the comparison of the model’s running efficiency,the new model took the shortest time to predict the test set 1000 times on all three datasets,compared to the slowest model,DUNet,that takes 5.6 hours,the model proposed in this paper only takes 9.5 minutes.Overall,the experimental results demonstrate that the new model proposed in this paper has good inference speed and segmentation accuracy.