Research On Image Segmentation And Lesion Grading Of Diabetic Retinal Vessels Based On Deep Learning

In recent years,the rate of diabetes is continuously on the rise,and a prevalent complication associated with this condition is diabetic retinopathy(DR).This medical condition has the potential to impair the retina and cause permanent blindness if it is not managed promptly.In the study of DR.retinal vessel segmentation and lesion grading are critical.Retinal vessel segmentation can help doctors determine the location and shape of retinal vessels,enabling accurate detection of DR’s impact on the fundus.Lesion grading helps doctors evaluate the severity of the lesions and develop the optimal treatment plan.Therefore,exploring an efficient and accurate method for retinal vessel segmentation and lesion grading has great practical value.With the outstanding performance of deep learning technology in various fields,significant progress has also been made in retinal vessel segmentation and DR lesion grading.However,the current retinal vessel segmentation has low accuracy and the lesion grading is not highly accurate.This article proposes a retinal vessel segmentation model and a DR lesion grading model based on deep learning.The research in this thesis focuses on the following two main aspects:(1)To address the problem of one issue in retinal vessel segmentation tasks is the poor segmentation accuracy for fine vessel branches and endpoints.this thesis proposes an improved UNet network-based retinal vessel segmentation model.Firstly,a swim transformer module is introduced into the UNet encoder.By employing the self-attention mechanism and leveraging the feature representation capability of the module,the model’s generalization ability and robustness have been significantly enhanced.Secondly,the skip-connection structure of the UNet network is redesigned,and a feature enhancement skip-connection module is used to reduce the sensitivity of the features to noise and interference.Finally,a multiscale feature fusion prediction module is added to fuse multiple sets of features of different scales.Test results show that the channel attention mechanism can enhance useful features and improve vessel segmentation accuracy.Testing the STARE and DRIVE public datasets,the accuracy,sensitivity,specificity,and AUC values of the proposed model has achieved 97.47%,80.87%,98.84%,97.38%and 96.72%,81.06%,98.22%,and 97.03%,respectively.(2)To address the issue of small differences between DR levels,the thesis proposes a transfer learning-based algorithm for grading diabetic retinopathy.The algorithm first uses a YOLOv5 object detection network to obtain DR-Related lesion areas and an improved UNet segmentation method to obtain retinal vessel images.Then,the lesion area features and retinal vessel features are merged with the original image by weighted fusion to obtain feature maps.Finally,the merged feature maps are input of an improved YOLOv5 classification network based on transfer learning to obtain the diabetic retinopathy grading results.Testing on the EyePACS public dataset showed that the model achieved an accuracy,precision,recall,and F1 score of 76.48%,77.50%,67.68%,and 77.08%,respectively,on evaluation indices.Furthermore,the accuracy of DR(0-4)grading has reached 78.68%,76.28%,68.86%,79.12%,and 80.45%,respectively.(3)The algorithm model proposed in Chapter 2 and Chapter 3 is the core algorithm of image processing.According to the demand analysis,the retinal blood vessel image segmentation and DR Classification visualization system are designed and realized.The two proposed methods have been significantly improved in DR retinal vessel segmentation and lesion grading through multiple sets of experiments comparison.