Research On Vessel Segmentation Methods Of Fundus Images Based On Deep Learning

The pathological changes such as sclerosis,exudation and hemorrhage of retinal blood vessels can reflect the changes of systemic health conditions.By assessing the condition of the retinal vessels,the physician can further understand the health of the patient’s body.In addition,the uniqueness of retinal vessels can also be used to build more secure recognition systems.Whether it is applied to disease diagnosis or building a certification system,the accurate segmentation of retinal blood vessels is crucial.Relying on ophthalmologists to segment color fundus images is inefficient and will consume a lot of financial and material resources.The rapid development of computer vision technology makes it possible to segment medical images automatically by computer.Compared with manual segmentation,the automatic segmentation algorithms based on computer vision are faster and more efficient.With the boom of artificial intelligence technology and the upgrading of computer hardware equipment,deep learning has developed rapidly and become the focus of various fields.Based on the above research background,this paper takes color fundus images as the research object to study the automatic retinal vascular segmentation approaches based on deep learning.The specific research results and contents are summarized as follows:（1）A segmentation network（DE-UNet）based on U-shaped model for fundus image segmentation is proposed.First of all,DE-UNet takes U-Net as the backbone.The symmetrical encoding-decoding structure and the introduction of skip connection can combine high-resolution information with low-resolution information,which is conducive to improving the accuracy of segmentation.Secondly,in order to make more efficient utilize of features and reduce the number of parameters,dense connection module is introduced into the network.Meanwhile,the channel attention mechanism is adopted in order to emphasize important features and suppress irrelevant features.In addition,the influence of the cross-entropy loss function and the hybrid loss function on the segmentation results is also discussed.Finally,experimental results on the benchmark datasets demonstrate the effectiveness of the proposed model.（2）A segmentation network（U~2R-GAN）based on generative adversarial network for fundus image segmentation is proposed.Generative adversarial network is a generative model consisting of generator and discriminator.Firstly,U~2R-GAN employs the lightweight U~2-Net as a generator to generate vessel probability maps.U~2-Net is composed of a series of U-shaped convolution module.This module can capture multi-scale features.It also allows to control the parameters of the network by adjusting the dimension of the feature maps in the module.Secondly,for the purpose of accelerating convergence and avoiding the degradation problem of the network,Res Net is applied as the discriminator to score the vessel probability maps from the generator.Generator and discriminator can improve their generative ability and discriminant ability respectively through the game.After a certain number of iterative trainings,higher quality vascular segmentation images can be extracted from the generator.Finally,the experimental results show that the proposed model can achieve good results in the color fundus image segmentation,and the sensitivity of the proposed model to the retinal blood vessel pixels has also been effectively improved.