Research On Retinal Blood Vessel Segmentation Method Of OCTA Images Based On Deep Learnin

Optical coherence tomography angiography(OCTA)has been widely used in clinical medicine due to its non-invasiveness and high resolution.Automatic segmentation of retinal vessels in OCTA plays an important role in the diagnosis and treatment of retinal diseases such as diabetes,glaucoma,and hypertension,and has certain clinical research value.However,most existing deep learning segmentation methods are based on region segmentation,and it is difficult to achieve accurate segmentation of retinal vessels with multiple and diverse branches in areas with no significant intensity differences.Therefore,accurately segmenting the complex and curvy structure of retinal vessels is an extremely challenging task that requires further optimization of its segmentation methods.In this context,this paper focuses on the retinal vessel segmentation task and proposes a deep learning-based OCTA image retinal vessel segmentation method,which includes the following aspects:(1)To address the problem of fine blood vessels being easily overlooked and incorrectly segmented in OCTA images,we propose an SS-Net based on a U-shaped network to achieve retinal vessel segmentation.In SS-Net,we introduce a module called SRes Block,which combines residual structures and split attention mechanisms to form the backbone of the encoder and decoder architecture.It effectively solves the problem of gradient disappearance and assigns greater weight to microvascular features.Additionally,SS-Net adds a spatial attention module to extract key information from the spatial dimension and focus on spatial features.SS-Net improves the attention to fine blood vessels from different dimensions,addressing problems such as discontinuous fine blood vessels being continuously segmented in previous networks.The accuracy,an important index,reached 0.9258/0.9377 in two subsets of ROSE with different depths dataset A and dataset B,and the Dice coefficient increased by about three percentage points compared with state-of-the-art segmentation models.(2)To address the problems of high noise,complex vascular structures,and intermittent segmentation of coarse vessels in OCTA images,we propose a DAS-Net based on Seg Net to achieve pixel-level segmentation of feature maps.DAS-Net encoder integrates deformable convolution blocks to capture the geometric transformation of vascular structures and adaptively adjust the receptive field based on the size and shape of retinal vessels.It better detects vascular structures of different scales and particularly enhances the detection of continuous coarse vessels.DAS-Net adds a global attention module to assign greater weight to key features in feature map channels and spatial dimensions,highlighting and enhancing the diverse vascular features,depicting the overall dependence of vessels,and also focusing on fine blood vessels.The effectiveness of different components of DAS-Net is verified by ablation experiments.Compared with various advanced network models,the experimental results on multiple datasets confirm the superiority of DAS-Net.The accuracy,an important index,reached 0.9278/0.9612/0.9480 in ROSE、OCTA-500＿3M、OCTA-500＿6M three different datasets,and the Dice coefficient increased by about two percentage points.