Retinal Vessel Segmentation Algorithm Based On Multi-scale Attention Network

Fundus disease is a major factor leading to irreversible blindness in humans and requires practical and effective measures to protect the eyes,such as carring out nationwide fundus screening campaign.However,the number of potential patients with retinal fundus diseases in China is relatively large,and it is difficult to diagnose them manually only through ophthalmic medical experts,and the number of doctors,technical level,and clinical experience all directly affect the accuracy of diagnosis.After the development of science and technology into the information era,the level and speed of computer data processing has been continuously improved,and the research on medical aspects using image processing and computer vision related technology is in full swing,especially in the field of medical imaging has a great development.Using the learning ability of computers to replace doctors to perform some work of quantifying and visualizing the relevant pathological structures,assisting or even replacing doctors to make accurate diagnosis and precise treatment of the disease,can achieve more effective patient assistance based on saving human and material resources.Therefore,there is great practical value in using deep learning techniques to implement vascular segmentation tasks,which are beneficial to the early diagnosis of fundus diseases.The proposed U-Net network based on the encoder and decoder structure largely solves the medical image segmentation problems such as retinal vessel segmentation.The structure with jump connections in U-Net can effectively connect the features of the encoder and decoder branches,but ignores the proportional changes within the path and the relevant regions in the feature map,while the proportion of vessels is small for retinal images,which makes the accuracy of network for blood vessels segmentation not high enough.To this end,this paper proposes a new multiscale attention network MA-Net for retinal vessel segmentation,with the following main contributions.(1)By improving the path scale variation problem in the U-Net network,the proposed multiscale network M-Net integrates the feature map information of each layer and utilizes multiscale connections at different levels to obtain more comprehensive information,learning information of both shallow and deep layers to ensure better feature representation.(2)The attention module is added to the improved multiscale network to suppress irrelevant regions in the feature maps and highlight salient feature information as a way to focus on the desired retinal vessel regions.(3)This paper also proposes a vessel width weight module,which is combined with the proposed multi-scale attention network MA-Net to improve the training loss function and make the network more effective for segmentation of fine vessels.The network is experimentally tested on the classical datasets DRIVE and CHASE_DB1 for retinal vessel segmentation.Due to the small number of vessel images in the dataset,the model is trained in this paper after the data volume is expanded by random cropping method.The experimental results show that MA-Net has numerical improvement in F1-score,accuracy AC and AUC of segmenting retinal vessels on both databases compared with previous methods such as U-Net,and the addition of vessel width weight module also make the segmentation of small vessels more precise.The F1-score on DRIVE and CHASE_DB1 datasets were 0.8183 and 0.7933,the accuracy was 0.9557 and 0.9594,respectively,and the ROC curve values reached 0.9677 and 0.9715.In general,the content studied in this paper had better performance on the retinal vessel segmentation task.