Research On Segmentation Algorithm Of Mouse Retina Image Based On U-Net Model

Diabetic retinopathy(DR)is a serious blindness.In the research and development of DR therapeutic drugs,the effect of drugs is often evaluated by observing the changes of pathological images of mouse retina under microscope,but this evaluation method is subjective and there is no quantitative standard.The area of nerve fiber layer(RNFL)and nuclear area in the pathological image of mouse retina can be used to evaluate the pathological changes of DR and the objective and provide a quantitative evaluation method of drug effect.Therefore,the segmentation of mouse retina image is of great significance for the evaluation of drug effect of DR.This thesis proposes a method to quantitatively evaluate the degree of DR by using deep learning algorithm.The improved U-Net model is used to realize the segmentation of mouse retinal image,which has been applied to the actual project of the development of DR therapeutic drugs and achieved good results.At the same time,due to the difficulty of pathological image labeling,small sample base,there are some problems such as poor generalization performance of the model,poor segmentation effect of the new mouse retinal image.Combined with the idea of transfer learning,this thesis uses two transfer methods to solve the problem of small sample size,and improves the segmentation performance of the model for the new mouse retinal image.The main contents of this thesis are as follows:1.Aiming at the problem of inaccurate RNFL segmentation of mouse retinal image,this thesis proposes a retinal image segmentation method based on improved U-Net model.First,the improved U-Net segmentation model is used to realize the RNFL segmentation in retinal image.Firstly,the residual path is used to reduce the information gap between features.Secondly,the Sub-pixel convolution layer is used to perform the upsampling operation to reduce the loss of detail information in the upsampling operation.Finally,the residual module is used to simplify the learning process,enhance the gradient propagation and reduce the model degradation caused by network depth.The experimental results show that the improved model can achieve better segmentation effect,and achieve RNFL basic accurate segmentation.Secondly,the segmentation of nucleus in retinal image is realized by threshold segmentation and Fusion Net network model.2.Aiming at the problem of quantitative evaluation of mouse retinal image,the segmentation results of retinal image are analyzed quantitatively,and the feasibility of the segmentation results for quantitative analysis of DR drug effect is verified.First of all,this thesis defines the relevant indicators with biological significance according to the actual changes.Secondly,the image segmentation results of mice retina with different age and dosage were analyzed.The results show that the quantitative analysis based on the segmentation results of retinal image is more sensitive to the change of DR drug effect than the human eye observation.The results show that the segmentation results of mouse retinal image can be used to quantitatively evaluate the effect of drugs.3.Aiming at the poor segmentation effect of the model for the new mouse retina image,this thesis combined with the idea of transfer learning to transfer the existing mouse retina image.In this thesis,we first use CycleGAN network and Reinhard algorithm to transfer the color and style of the retinal image,and generate the image with the new style and color of mouse retinal image.Then we add the transferred image to the original image data set for training.The experimental results show that the new model can improve the segmentation effect of the new mouse retina image and reduce the need of image annotation.