The Study Of Fundus Retinal Vessel Segmentation Methods For AI-aided Diagnosis

Fundus images contain rich morphological information,and the incidence of blinding eye diseases is highly correlated with changes in retinal morphology,which is clinically detected by segmenting retinal structures in fundus images.Deep convolutional neural network(DCNN)-based methods have achieved great success in the automatic segmentation of retinal vessels.However,the task of fundus retinal vessel segmentation also faces great challenges.On the one hand,noise is introduced when fundus images are acquired,and the image contrast is low with unclear vascular structures; on the other hand,the structure of fundus images is complex,and lesions such as exudate may be incorrectly segmented as vascular pixels.To solve these problems mentioned above,this paper aims to develop an AI-aided diagnosis system for the segmentation of retinal vessels in fundus images.The main work and innovation points of this paper are as follows:(1)To address the problem of insufficient detail information capture capability,a multitask symmetrical network is proposed to extract global features and detail features respectively through symmetric feature extraction branches generating different size receptive fields.A fusion network merges the features to output a segmentation result that combines both global and detailed features,improving segmentation accuracy of retinal vessels.(2)To address the problem of large number of parameters in multi-task symmetrical network,a single-path multi-scale attention-guided fusion network is proposed.In order to make full use of channel information from deep layers and spatial information from shallow layers,an attention-guided fusion block is designed to fuse the features from different network layers,which significantly reduces the number of parameters.(3)To address the problem of low fusion efficiency in the multiscale attention-guided fusion network,a dual-path progressive fusion network is proposed.It includes a convolutional path for detecting local features and a recurrent convolutional path for extracting contextual information,enabling the acquisition of sufficient detailed information and rich contextual information.The proposed progressive fusion strategy promotes feature flow at different levels by fusing features from the same scale,adjacent scales,and all scales,improving feature fusion efficiency.(4)To address the issue of a large amount of lost detailed information in the convolutional and pooling processes,an edge dimension attention-enhanced network is proposed.It enhances both local edge information and global semantic information,and promotes enhanced deep features flowing to shallow networks.The prediction outputs from different decoding stages are learned by an adaptive weight learner to obtain the best weight combination and weighted output of the final segmentation result.The network enhances the feature representation in the encoding stage.Extensive experiments are conducted on three typical publicly available retinal vessel segmentation datasets.Among them,ablation experiments are introduced to validate the effectiveness of the proposed network modules,the advancedness of the proposed network is verified by comparison experiments with other segmentation methods,heat map and visualization results are provided to visualize the performance ofthe proposed segmentation method,and complexity experiments are also included to balance the segmentation accuracy and inference speed.