Research On Retinal Blood Vessel Segmentation Based On Fully Convolutionalneural Networks

It is helpful for doctors to locate human eye diseases and provide objective and accurate analysis results by automatic retinal blood vessel segmentation technology.There are differences in the shooting equipment,shooting angle,and lighting environment during the acquisition of the fundus retinal image,which leads to the problem that the overall image tends to be red and the contrast between the blood vessel and the background is low.Retinopathy symptoms has unexpected colors and morphology,making it more difficult to segment blood vessels from background noise.Although many researchers have developed some solutions to image segmentation tasks based on different technologies,the proposed solutions still have problems such as low accuracy,inaccurate capillary segmentation,and sensitivity to lesions.Aiming at the problems of fundus vascular segmentation tasks and existing methods,this paper designs a novel retinal vascular segmentation framework based on the Fully Convolutional Neural Networks.The innovations are as follows:(1)In order to deal with the problems of shooting equipment,shooting angle,lighting environment,and image color contrast during the acquisition process,this paper explores four-step preprocessing strategies for retinal image preprocessing to reduce the impact of these problems.The four-step preprocessing strategy consists of grayscale processing,data normalization,Contrast Limited Adaptive Histogram Equalization,and Gamma Nonlinearity.From the image processing results,this four-step strategies can improve the discrimination between blood vessels and background in the image,which can better allow the model to learn the feature differences between blood vessels and background and perform accurate segmentation.The experimental results show that the four-step preprocessing strategies can improve the performance of the trained model in multiple evaluation indicators.(2)To make full use of a limited number of retinal images,this experiment uses a dynamic patch extraction algorithm to dynamically extract small batches of patches during training.A new data augmentation method called Random Local Replacement is proposed to further utilize the limited retinal image to provide more sufficient data for the training model.The experimental results show that the Random Local Replacement method can improve the segmentation performance of the model and obtain better experimental results after combining with the conventional data augmentation algorithm.(3)This paper proposes two improved structures based on the Fully Convolutional Neural Network: the first is called Basic Fully Convolutional Neural Network(BFCN);the second is called Multi-scale,Multi-path,and Multi-output fusion Fully Convolutional Neural Network(M3FCN).BFCN is designed for the task of retinal image segmentation,and M3 FCN is an improved model for the deficiency of BFCN.M3 FCN has better structure than BFCN by constructing multi-scale input layer,multi-path structure and multi-output fusion module.M3 FCN has a deeper network structure,which increases the capacity of the model,and skip connection with different jumping distances greatly reduces the difficulty of training the model.Comparative experimental results show that multi-scale input layer,multi-path structure,and multi-output fusion module can make M3 FCN perform better in segmentation tasks.Finally,the proposed framework was trained and tested using three retinal image datasets collected under different equipment and different acquisition conditions.After comparing the results of F1 score,accuracy,and other indicators,we can see that M3 FCN has reached competitive experimental results on the three datasets of DRIVE,STARE and CHASE.The experimental results obtained by M3 FCN in the three data sets are as follows: F1 scores are 0.8321,0.8531,and 0.8243,and the accuracy rates are 0.9706,0.9777,and 0.9773,respectively.