Research On Retinal Vascular Image Segmentation And Lesion Detection Based On Deep Learnin

Diabetes retinopathy(DR)is an eye disease that endangers vision.Early diagnosis and treatment are very important for DR patients.In clinical diagnosis,ophthalmologists usually segment and detect DR based on fundus images captured by fundus cameras.The entire process depends entirely on the experience of ophthalmologists,which is both timeconsuming and labor-intensive.With the development of big data and deep learning,computer-aided medical diagnosis has become a research hotspot,but for medical images,data acquisition is often limited.Based on this,the research on retinal vascular image segmentation and DR image lesion detection has the following problems:(1)there is a lack of large publicly annotated DR image datasets,and some datasets have imbalanced data;(2)The mainstream retinal vascular image segmentation methods have the problem of difficult segmentation of vascular endings and easy loss of details.In response to the above issues,this article conducted in-depth research in the following two aspects.(1)Propose a retinal vascular segmentation model RCRAMU Net based on U-Net network,Recurrent Convolution(RC),Residual Block(RB),and Attention Mechanism(AM).The Accuracy of CHASEDB1 and DRIVE in the public dataset were 96.92% and 96.95%,respectively,with sensitivity of 81.42% and 83.42%,and specificity of 99.07% and 98.31%,respectively.The experimental results show that the RCRAMU Net model can effectively improve the segmentation effect of retinal vascular images’ vascular endings,preserve certain vascular details,and have a certain improvement compared to existing algorithms.(2)We propose a DR fundus image five classification model EB3 MTNet that integrates Efficient Net network,Multi Task Learning(MT),and Focal Loss function.We conducted indepth research on training,validation,and testing on the DR2015 dataset.The experimental results show that,in the case of imbalanced data,the EB3 MTNet model achieves a classification Accuracy of 96.7%,specificity of 98.6%,sensitivity of 94.9%,and a quadratic weighted kappa score of 92.5%,with improvements of 1.4%,0.2%,1.2%,and 1.74%,respectively.It has good classification performance in DR image five classification tasks.This study effectively addresses the shortcomings of current research on DR image segmentation and lesion detection.The retinal vascular segmentation method based on the RCRAMU Net model not only improves the segmentation Accuracy of the network in vascular endings,but also alleviates the noise problem generated in low contrast retinal vascular images.Therefore,the RCRAMU Net model can better provide clinical diagnostic references for ophthalmologists;The research on DR lesion detection based on EB3 MTNet model achieves five classifications of DR images,effectively solving the problem of imbalanced distribution of DR image categories.At the same time,multi task learning is introduced to make the model consider the cost of misclassifying DR levels,alleviating the diagnostic pressure of ophthalmologists.In summary,the research on retinal vascular image segmentation and DR image lesion detection is of great clinical significance,which helps doctors diagnose ophthalmic conditions.