Research On Retinal Fundus Image Segmentation Based On U-Net And GAN

Retinal fundus image segmentation has important clinical application value.However,there are still many deficiencies in the current segmentation method based on deep learning,such as the segmentation effect of micro blood vessels is not good enough,blood vessels are prone to fracture and loss,and the accuracy and sensitivity of segmentation need to be further improved.Considering the above shortcomings,this thesis proposes a new segmentation network AMSFU-Net.Compared with other methods based on U-Net,the segmentation quality of micro blood vessel has been greatly improved.On the other hand,we use the proposed AMSFU-Net as Generator,and design a new generative adversarial strategy and loss function for the Generator and Discriminator,proposing a new GAN model named Retinal-GAN,which surpasses the performance of many GAN based methods in fundus vascular segmentation,and we further improve the robustness of AMSFU-Net model through adversarial learning.The main contributions of this thesis are as follows:(1)In AMSFU-Net,we propose two kinds of Attention modules in Encoder and Decoder stages,adjusting features weights according to the importance of features,and realizing the recalibration of features,so that the model focuses on the important areas and positions of vascular feature map,improving the accuracy of model segmentation.(2)We design a multi-scale feature extraction and aggregation module.With the combination of dilated convolution with different expansion rates and different kernel sizes,we obtain a variety of multi-scale features with different receptive field sizes,which is helpful for the model to identify the vascular structures with different shapes and thickness.(3)In the Encoder path of AMSFU-Net,we patch embedding the feature maps of different Encoder stages,obtain the token sequences of multi-scale features,and establish the global context information dependency of the multi-scale feature sequences through the Transformer Encoder,so as to improve the segmentation ability of the model for tiny and complex blood vessels.(4)We take AMSFU-Net proposed based on innovation points(1),(2)and(3)as the Generator,and design a two-classification Discriminator simultaneously.Through the generative adversarial strategy based on image pairs,we design a new form of loss function and training strategy for the Generator and Discriminator respectively,and propose RetinalGAN model,which improves the quality and criterion of vascular segmentation of the traditional GAN method,the segmented micro-vessels appear clearer and more continuous.From the experimental results,on DRIVE dataset,the accuracy and sensitivity of our AMSFU-Net are 96.26% and 83.92% respectively,and 97.30% and 81.24% on STARE dataset.The accuracy and sensitivity of our Retinal-GAN model reach 96.60% and 84.11%on DRIVE and 97.43% and 83.22% on STARE.In general,our AMSFU-Net and RetinalGAN have higher segmentation performance indexes than other methods based on U-Net and GAN.The segmented blood vessels are relatively less missing and broken,the segmented blood vessels are clearer,and the segmentation effect of some relatively complex blood vessels such as micro blood vessel is also better.