Research On Retinal Vascular Image Segmentation Based On U-Net

The eye is not only a window to the soul,but also an effective way to observe the deeper vasculature of human tissues.The geometry of fundus vascular structures contains rich focal information,which can provide an important basis for the screening and diagnosis of many fundus diseases.Due to the harsh imaging conditions and large data volume of fundus images,large-scale vessel segmentation cannot be achieved by manual segmentation.Therefore,automated vascular segmentation algorithms are of great research value.In this thesis we investigate the vascular segmentation technology of retinal images based on deep learning to achieve automated and high-precision vascular segmentation.The main research contents of this thesis are as follows.1.In the vascular segmentation work of retinal images,this thesis proposes a U-Net-based vascular segmentation method for retinal images: FSU-Net.through multi-scale pooling and full-size feature fusion,low-level texture features and high-level semantic features are fused to expand the sensory field while fully incorporating contextual information to improve vascular segmentation capability.Meanwhile,residual learning is introduced to make the network converge faster.Experiments show that the algorithm in this thesis can better preserve the image texture features compared with the mainstream retinal vessel segmentation algorithms and can achieve accurate vessel segmentation.2.In order to further improve the network’s ability to extract vascular features,improve the expression of effective features such as vascular structure and suppress the expression of invalid features such as background noise,an attention mechanism is introduced,and a retinal image vascular segmentation method based on the attention mechanism and U-Net is proposed: FSAU-Net.this method uses the attention mechanism to learn the effectiveness of features when full-scale features are fused,and uses the method uses less a priori knowledge to obtain the most relevant blood vessel features to achieve accurate blood vessel segmentation.Experiments show that the algorithm in this thesis can effectively suppress background noise and has better vessel segmentation ability compared with mainstream retinal vessel segmentation algorithms.Based on the U-Net model,this thesis proposes two improved retinal vascular segmentation networks: FSU-Net and FSAU-Net,which can effectively segment retinal images,and the segmentation effect is better than the traditional unsupervised algorithm and supervised algorithm.