Research On Segmentation Algorithm Of Retinal Blood Vessels Based On Deep Learning

With the development of computer technology,methods represented by deep learning have opened up new directions for the research and development of medical images.Due to the important connection between retinal blood vessels and diabetes such as hypertension,cardiovascular and cerebrovascular diseases,researchers have paid great attention to the medical task of automatic segmentation of retinal vessels.At present,in most retinal vessel segmentation methods based on deep learning,the encoder-decoder segmentation model integrates the global and local information of the retinal image that has made a breakthrough in the segmentation performance.However,due to the complicated morphological changes of retinal blood vessels,the influence of the lesion areas and the many and fine branches of capillaries,the previous segmentation algorithms have problems of over-segmentation and under-segmentation of retinal vessels,especially capillaries.At the same time,the simple improvement of the basic segmentation model lacks medical explanation.Therefore,on the basis of existing research methods,the retinal blood vessel segmentation algorithm based on deep learning has been emphatically studied in this paper,mainly including the following three aspects:First of all,in order to solve the problem that the deep neural network reduces the resolution of the image through continuous convolution and pooling operations resulting in the loss of detail information on the edge of retinal vessels and the discontinuous segmentation of small vessels,a full-resolution densely connected network is designed to segment retinal vessels.The continuous dense connection block structure enhances the network's ability to represent the features of blood vessels,which enables the network to fully extract the rich contextual semantic information and key detailed feature information of blood vessels,so as to improves the sensitivity of the network to blood vessels with similar texture structures.The expanded convolutional layer of the mixed expansion domain is used to increase the receptive field of the network feature map to obtain the context information of blood vessels in different scales,which enhances the diversity of blood vessel characteristics in the network.Experiments on the DRIVE and STARE datasets respectively show that the proposed network can effectively improve the performance of the basic segmentation network.Secondly,As the network depth increases,the background noise in the encoder is transmitted to the decoder through skip connections and the background pixels of the lesion area are easily segmented into vessel pixels by mistake.Considering the difficulty of segmentation of retinal vessel under the influence of the lesion area,a new multi-scale channel attention network is proposed.In the coding part of the deep layer of the network,the multi-scale module is used to capture the multi-scale features of retinal vessels to improve the feature extraction ability of the network.In the skip connection part,the channel attention module is merged to suppress redundant features and weaken the influence of the segmentation disconnection caused by noise information which enhances the ability of the network to model the blood vessels in the lesion area.The redesigned decoding block not only retains more semantic information,but also helps to optimize the training of the network better.The designed experiment obtained good segmentation results in two public datasets and effectively improved the segmentation effect of capillaries in the lesion area.Finally,based on the complex morphological changes of retinal vessels,the characteristics of the scale and structural changes of vessel branches in different regions are of key significance to retinal vessel segmentation.This paper proposes an efficient retinal vessel segmentation network with adaptive morphological changes.The network designs an adaptive morphological change regularization convolution module which uses deformable convolution to adaptively acquire the rich geometric deformation characteristics of blood vessels in the image,so as to enhance the modeling ability of the network for the morphological changes of the vessels and improve the generalization performance of the model.In addition,the regularized convolutional layer is used to achieve a stronger representation of the vessel features of the model which accelerate the training efficiency of the network and reduces the problem of overfitting.Experiments have proved that the network proposed in this paper reduces the influence of complex morphology on the segmentation performance of retinal vessels,and effectively improves the segmentation effect of the branch structure of blood vessels.