Research On Retinal Vascular Image Segmentation And Hypertension Quantification Method Based On Deep Learning

Fundus retinal vessels play a crucial role in noninvasively assessing cardiovascular diseases in humans.The morphological structure of these vessels reflects the severity of such diseases to some extent.Utilizing deep learning,this research paper focuses on retinal vascular image segmentation and hypertension quantification.The specific contributions are as follows:(1)A novel subsampling method called pixel fusion pooling(PF-Pooling)is proposed to address the issue of information loss during down sampling in traditional convolutional networks.This method splits and stacks the original feature images using parity difference and extracts key information through 1×1 convolution.Experimental results on different datasets demonstrate improved performance: an increase of 0.23% and 0.03% in F1-score,0.36% and1.98% in accuracy rate,and 0.23% and 3.84% in sensitivity for the DRIVE and STARE datasets.(2)To enhance the segmentation of retinal tiny blood vessels,a new algorithm is introduced by fusing images from different feature layers within the U-Net framework.In the decoding stage,fusion of information from various layers preserves more details.Additionally,an attention mechanism is incorporated into the convolution process to improve the model’s accuracy in segmenting tiny blood vessels.Experimental results on a public dataset indicate notable improvements over the original algorithm: a 0.47% increase in accuracy,0.7% increase in AUC,and 1.78% increase in F1-score for the DRIVE dataset.Moreover,the proposed algorithm achieves a segmentation accuracy of 96.67%,F1-score of 83.82%,and AUC index of 98.90% on the CHASE＿DB1 dataset.(3)In order to assist doctors in diagnosing hypertensive patients,a method for quantifying changes in retinal blood vessel morphology is presented.This involves extracting regions of interest(ROI)from segmented vascular images,performing linear fitting on the blood vessels in the region,and calculating the forward included angle of the fitted vascular curves.Experimental results demonstrate the effectiveness of the method,showing good robustness to image rotation and scaling.When compared with manual annotation data from professional doctors,the proposed algorithm exhibits an average error of only 0.69°,serving as a valuable auxiliary tool in diagnosis.Overall,this research contributes to the field of retinal vascular analysis by introducing novel methods for segmentation,fusion,and quantification,ultimately aiding doctors in diagnosing cardiovascular diseases and improving patient care.