Research On Retinal Vessel Image Segmentation Algorithm Based On Deep Learning

Retinal vessels are the only clear vessels in the cardiovascular system that can be observed in a non-invasive and non-traumatic manner.The precise segmentation of retinal vessel images has important clinical value for assisting doctors in the diagnosis and treatment of diseases such as retinopathy and hypertension.The traditional artificial segmentation method is not only time-consuming and laborious,but also has low segmentation accuracy.With the development and application of deep learning technology in the medical field,researchers have proposed many vascular image segmentation methods and obtained certain segmentation effects.However,retinal vascular image segmentation still has some problems such as low segmentation accuracy and easy rupture of small vessels.In order to solve this problem,two different segmentation algorithms are proposed in this paper,and experiments are carried out on foreign public data sets respectively,and a better retinal vessel image segmentation effect is obtained.The main innovations of this dissertation are as follows.(1)In order to solve the problems of the limited number of paths to obtain the information flow of the image in the U-Net network model and the difficulty in segmentation of small retinal blood vessels,a multipath U-Net based retinal blood vessel image segmentation algorithm was proposed.First,when the retinal blood vessel image is segmented,more retinal blood vessel information flow paths are obtained through the improved U-Net network model,and more retinal vessel characteristic information can be obtained.Secondly,the ordinary convolution in the improved U-Net network model is replaced with dilated convolution,which can expand the network’s perception horizon without increasing network parameters.The experimental results on the DRIVE public datasets show that the average accuracy,sensitivity and specificity of the model reached 95.64%,80.37% and 98.14%,respectively.The area under the receiver’s operating characteristic curve reached 0.9802,which is higher than the traditional segmentation method and achieved good segmentation results.(2)Aiming at the problems of low precision of retinal vessel image segmentation,easy rupture of small vessels and poor anti-interference ability of the model,a new retinal vessel image segmentation algorithm based on generating adversarial network was proposed.The algorithm mainly introduces an improved U-shaped network structure in the generator module that generating adversarial network.First,theresidual-dense module is added to the downsampling layer of the original U-shaped network structure,which can make full use of the image feature information of the multi-layer network,enhance feature propagation,and reduce the network model parameters.Secondly,on the original U-shaped network structure.The upsampling layer adds a attention mechanism module to focus feature information on blood vessels of different sizes in the retinal vessel image.The experimental results on the DRIVE and STARE public datasets show that the average segmentation accuracy of the model is 96.01% and 97.52%,respectively,and the sensitivity is 82.37% and 83.67%,respectively.The specificity is 98.79% and 98.64%,respectively.The area under the receiver’s operating characteristic curve is 0.9892 and 0.9899,respectively.Experiments show that this model has better segmentation accuracy and robustness than other models.