Research On Algorithm Of Retinal Vessels Segmentation Based On Deep Learning

Retinal blood vessel segmentation has always been a hot research task in medical image segmentation research.The extraction of blood vessels in the fundus image can tell whether a person has certain diseases,as well as the severity of some diseases like diabetes and arteriosclerosis.In practice application,it can not only solve a large number of manual segmentation tasks,relieve the pressure of manual operation,the shortage of human resources,and the pressure on doctors,allow doctors to spend more time and energy on disease consultation,but also promote the acceleration of intelligent medical treatment.In this paper,the task of retinal vessel segmentation was deeply studied.To the problem of vessel segmentation in fundus images,the image preprocessing and the vessel segmentation model based on deep learning was firstly introduced,different image enhancing techniques were applied,then the vessel segmentation network was improved through the feature analysis of the fundus vessels,finally the vessel segmentation model Retina-GAN was proposed.The experiments showed that the Retina-GAN model showed good accuracy on public datasets.The main research work of this paper is as follows:（1）Pretreatment of fundus images.Aiming at solving the problem of poor contrast and small color range of fundus images,the related image enhancement technology was studied,and an image enhancement method combining automatic color equalization algorithm and Gamma correction was proposed.Firstly,automatic color equalization was applied to adjust the brightness and the color of the fundus images so that the fundus images could have richer color and higher contrast.Then Gamma correction was used to convert the fundus images to gray ones.The above two steps made the boundary between the blood vessels and the background clearer.Furthermore,taking into account the small number of images in the dataset,the dataset was expanded by means of flipping,multi-angle rotation and random splicing after equipartition.（2）A new retina vessel segmentation model Retina-GAN based on Generative Adversarial Networks（GAN）was proposed.After in-depth study of the current domestic and foreign research situation of the retina segmentation,the paper found that the main problem of the current vessel segmentation model was that it tended to ignore the pixel points in the microscopic vessel edges.In this paper,a Retina-GAN vessel segmentation model was proposed by virtue of GAN’s better optimization effect to the generator through its iterative training and R2U-Net’s relatively better effect on segmentation details.Retina-GAN made use of R2U-Net as the generator of the network of GAN.To expand the receptive field and transfer parameters,R2U-Net was improved as follows.Firstly,a soft-attention unit dilated convolution was applied at the bottom of the network,secondly,the attention mechanism based on threshold segmentation was added into the network,which made the attention area of the model training more accurate through the method of mask extracting.To reduce over-fitting,the global average pooling of attention units based on convolution channel is applied in the discriminator.By virtue of the adversarial game principle in GAN,the Binary Cross-Entropy was used to iteratively train the model network.Experiments verified that the model proposed in this paper gained better segmentation results on the general evaluation criteria,which brought certain research value for retinal vessels segmentation.（3）With the principle of simple operation and good effects in software designing,a Web application system of vessel segmentation in fundus images was designed and implemented.The vessel segmentation system was mainly divided into three modules: the user login module,the image segmentation module and the result display module.The web page was simple,clear and easy to operate.The retina vessel segmentation model was applied in the image segmentation module of the software,which is a practical application in terms of the research result of this thesis.