Research On Segmentation Algorithm Of Retinal Vessel Image Based On Deep Learning

Automatic segmentation of retinal blood vessel images is a research hotspot at domestic and international.The characteristic changes such as the diameter and wall of the retinal blood vessels are closely related to the condition of the blood vessels throughout the body,which can assist in the diagnosis and prediction of cardiovascular diseases,diabetes,glaucoma,retinal vein occlusion and other diseases.Traditional extraction algorithms have the problem of inaccurate segmentation in the case of lesions,detailed areas,and uneven illumination.Aiming at the main defect of low segmentation accuracy of lesion area,this paper proposed two methods to improve the network structure based on deep learning for retinal image blood vessel segmentation,which can effectively improve the segmentation accuracy.The main research contents are as follows:(1)In order to solve the problem of uneven illumination and low color contrast in retinal blood vessel images,firstly,the data set is first preprocessed and then input to the network for training.The preprocessing steps are accordingly:channel fusion,normalization,contrast-limited adaptive histogram equalization and gamma correction.Image preprocessing can improve the contrast between the blood vessel and the background,which is helpful for the model to learn the characteristic information of the blood vessel and improve the accuracy of segmentation.(2)An improved retinal vessel extraction algorithm combining U-Net and AC-Net is proposed.Aiming at the problem of poor segmentation of the original U-Net in the lesion area,the U-Net deep neural network was improved.The dense connection method of Dense-Net is added at the bottom of the network to realize feature reuse and avoid over-fitting caused by continuous convolution of a series of convolutional layers.In this paper,the characteristic information of BCONVLSTM combined encoder and decoder,combined with the idea of AC-NET,proposed MultiAC module,and added to the process of U-NET down sampling and up sampling,to help the network to learn more complex characteristic information.The segmentation accuracy of the improved network is improved,and better segmentation results are obtained in the focal area.(3)The paper proposes a residual U-Net algorithm that combines multi-scale dilated convolution and attention model.Aiming at the long training time of the improved algorithm in(2),the above algorithm is improved to reduce the time cost.First,the original U-Net is used as the baseline and combined with ResNet,and the residual structure of ResNet is added to U-Net,which can effectively solve the problem of gradient disappearance and improve the efficiency of network training.Secondly,the attention mechanism is introduced in the ResU-Net jump connection to make the network model more focused on learning vascular features,reduce the interference of factors such as background and lesions,and improve the training speed and segmentation accuracy of the model.We consider the burden of training time brought by the depth of the network,reduce the number of layers in the network structure,and introduce a multi-scale hole convolution strategy at the bottom of the network to increase the receptive field.This paper solves the problem of a few feature information loss during the downsampling process.The ablation experiment of the above improved algorithm verifies the effectiveness of each module and improves the performance of the network.The image segmentation results of the improved algorithm experiment and the objective quantitative index data show that the improved algorithm has good segmentation performance in the case of lesion areas and complex backgrounds.Compared with the first improved algorithm BDAU-Net,the second improved algorithm MDARU-Net has good time performance.