Research On Retinal Vessel Segmentation Algorithm In Fundus Images

Objective: Diabetic retinopathy is closely related to changes in the width,angle,and branch shape of the fundus retinal blood vessels.Changes in fundus retinal blood vessel characteristics can reflect diabetic retinopathy and can also help doctors analyze the disease and diagnose and treat patients.Therefore,a retinal vascular segmentation algorithm that meets the needs of clinical testing,has high accuracy and fast operation speed is of great significance for early diabetic retinopathy screening and later treatment.Now that digital image processing technology has been widely used in the medical field,segmentation of retinal blood vessels on computer systems has become a new type of medical research.Method: Method 1 is based on Frangi and Ostu’s retinal vascular segmentation algorithm.First,the original color retina image is preprocessed,and then the blood vessel image is edge-detected based on the Frangi filter,and finally the segmentation result image is obtained;the pre-processed and enhanced retinal image is segmented using Ostu threshold segmentation.Finally,the two result images obtained by the two algorithms are fused to obtain the final image.Method two is a retinal vessel segmentation algorithm based on morphological transformation and multi-scale line detection.First pre-process the image,then enhance the pre-processed image,and finally use multi-scale linear detection to extract blood vessels.Method 3: Retina vessel segmentation algorithm based on Attention U-net.First,the grayscale,standardization,and contrast-contrast adaptive histogram equalization(CLAHE)of the image in the DRIVE database are followed by gamma transformation.In order to prevent overfitting,the data volume is increased.Construct Attention U-net retinal vessel segmentation network,send the pre-processed training set to Attention U-net network training network,and finally send the test set image to the trained retinal vessel segmentation model to obtain segmentation results.Results: Method 1: The average sensitivity of the DRIVE and STARE data sets was 0.5811 and 0.7866,specificity was 0.9311 and 0.9639,and accuracy was0.9383 and 0.9581.Method 2: The sensitivity on the DRIVE and STARE data sets is 0.7670 and 0.7586;the specificity is 0.9520 and 0.9556;the accuracy is0.9590 and 0.9611.Method 3: The sensitivity on the DRIVE data test set is0.7758,the specificity is 0.9769,and the accuracy rate is 0.9668.Conclusion: Method one is simple and efficient,which can suppress noise well,but it is susceptible to interference from non-vascular vessels such as optic disc.The second method is robust,and the segmentation speed is fast,which solves the interference of non-vasculars such as the optic disc,but the microvascular segmentation at the end of the blood vessel is not ideal.Method3 has good connectivity and integrity to the microvessels at the end of the blood vessels,and it has achieved good results in segmenting the microvessels.