Research On Retinal Vessel Segmentation Based On Deep Learning

As the only observable terminal vessel in our body,the morphological characteristics of retinal vessels are important biological markers.Professional physicians can use this marker to accurately diagnose the type and severity of patients’ disease.Morphological information of retinal vessels can reflect clinicopathological features.The current method of obtaining retinal vascular morphology information relies on manual segmentation.Accurate retinal vascular segmentation results are of great significance for early diagnosis of patients.However,the contrast of retinal images is insufficient and the shape of vessel is complex,which will reduce the accuracy of retinal vessels segmentation.More importantly,manual segmentation of retinal vessels is not only inefficient but also susceptible to subjective factors from doctors.To address the aforementioned issue,this article proposes a deep learning-based retinal vessel segmentation method that can accurately recognize and extract retinal vessel features,completing the task of automatic retinal vessel segmentation.First,in order to enhance image contrast,reduce noise in the data,and meet the needs of deep learning,the preprocessing operations of image enhancement,image noise reduction,feature scaling,and data enhancement are performed on the dataset.Subsequently,a retinal vessel segmentation algorithm is designed,which uses Residual Attention with Identity(RCI)as the infrastructure architecture of semantic segmentation network feature extraction,and increases the weight of vessel with lower resolution to improve the accuracy of retinal vessel segmentation;Atrous Spatial Pyramid Pooling with Deformable Convolution(ASPPD)enables the network to acquire a larger receptive field and adaptively learn the tree-shaped vessel morphological characteristics;Asymmetric Convolution Edge Refinement(ACER)is used to improve the fit and refinement of retinal vessel edges so as to enhance the segmentation effect of the retinal vessel.Finally,multiple sets of experiments were conducted to verify the effectiveness and generalization of the proposed retinal vessel segmentation algorithm in this paper.The performance indicators of the retinal vessel segmentation method proposed in this paper tested on the OCTA-500 dataset,including Dice,Jaccard,Binary Accuracy,Accuracy,and Precision,were 89.35±2.63%,80.84±4.05%,93.86±1.47%,98.06±0.67%,and 90.15±4.18%,respectively.Compared with other segmentation methods,the method proposed in this article has better segmentation performance for retinal vessels,and this improvement is not limited to a single dataset,but also applies to other publicly available datasets.In the predicted retinal vessel images,the retinal vessel edges are smooth,and the over-segmented and under-segmented regions are minimal.It shows that the comprehensive performance index of the retinal vessel segmentation algorithm proposed in this paper is superior to other methods with certain clinical application values.