Research On Classification Task Of Retinopathy Screening For Fundus Images Based On Attention Mechanism

The eye is the main organ of the human body that receives information from the outside world,and the health of the eye is very important to the working and living of human beings.The majority of retinopathy is caused by retinal lesions in the fundus.Meanwhile,machine learning technology has been deeply participated in the field of fundus image diagnosis,but researchers also face many challenges: the similarity of images in specialized fields requires models to capture more subtle features;medical datasets have serious class imbalance problems;and the diagnosis of diseases requires specialized knowledge,and how to incorporate this expertise into deep models is a problem that researchers need to consider.Based on the above problems and challenges,this paper proposes an attention-guided retinopathy screening network based on fundus images,and the main work and innovation points are as follows.(1)For AMD diseases,the corresponding vessel maps are obtained using pretrained segmentation models,and the dynamic attention module is designed to calculate the corresponding spatial attention and combine the feature maps at different levels in the classification model,respectively.Comparison and analysis for the validity of global images,optic cup regions,and vessel regions and ensemble methods to fuse the information of different regional streams are also included,which in the end increase result of AMD screening from 0.812 to 0.846 in term of AUC.(2)Extending our network to the multi-disease classification task,the attention module was redesigned considering fusing spatial attention and channel attention and bi-directional propagation of global information and regional information.Furthermore,the multi-label loss function was modified to further improve the disease diagnosis by considering the class imbalance from dataset.As a result,multi-disease classification criteria weightedAUC is improved from 0.736 to 0.812.In addition,the use of attention map can increase the model interpretability and help users to analyze cases.