Research On Fundus Retinal Segmentation Based On Deep Learning

Eye diseases such as diabetic retinopathy and diabetic macular degeneration are the leading cause of blindness worldwide.Diabetic macular disease has a long pre-clinical stage in which vision is not affected,and once a patient visits an eye clinic,vision loss is often irreversible.Early detection can prevent disease development and protect eyesight.Therefore,studying and analyzing the geometrical characteristics of retinal blood vessels,such as vessel diameter,branch angle and branch length,has become the basis of medical application for early diagnosis and effective monitoring of retinopathy.Retinal image assessment by an ophthalmologist is an important step in retinal pathology identification.Deep learning is applied to the retinal vascular segmentation of the fundus to improve the accuracy of vascular segmentation,which has important research significance and application value for the diagnosis and monitoring of eye diseases,prevention of disease development and protection of vision.Aiming at the main problems of retinal segmentation,the deep learning-based retinal vascular segmentation method was studied.Aiming at the problems of high illumination unbalance,color imbalance and low contrast of blood vessels in the color retinal images,the image was enhanced and optimized by grayscale,adaptive histogram equalization with limited contrast,gamma correction and standard normalized image preprocessing.In order to reduce the overfitting and enhance the generalization ability of the model,the data enhancement methods of random flip,rotation,affine transformation and elastic transformation are adopted to increase the number of data sets.By designing a convolutional neural network model,a segmentation model based on attention mechanism and network cascade is proposed.Based on the Unet network model,the attention-gated mechanism is applied to the upsampling in the decoder to selectively extract image features by identifying significant areas of the image and pruning characteristic response parameters.At the same time,a cascade network is adopted to realize the jumping connection between networks to compensate for the information loss during the feature map sampling.Correct errors in the vascular probability graph generated by a single network.Experimental comparison was made through ablation experiments.The results show that the segmentation effect is better in different data sets,and it is suitable for the segmentation of retinal vascular images.Figure 37 Table 4;Reference 50.