Research On Diagnosis Of Retinopathy Based On Self-Attention Learning

Nowadays,the health of the eye is of increasing concern,and fundus images are the most frequently used medical images in fundus disease detection,which include not only structures such as optic cup,optic disc and blood light,but also some lesions that arise when disease is present,such as exudates and micro aneurysms.Convolutional neural networks have made some achievements in the field of retinopathy-assisted diagnosis.However,due to its local operation,it is difficult to retain the global contexts information and long dependencies of fundus images,resulting in general performance when dealing with complex lesions or tissue structures in fundus images.This thesis is based on selfattention learning about retinopathy-assisted diagnostic techniques.The main work is as follows.Research on retinal disease detection method based on non-local attention learning.To address the problems of traditional convolutional neural networks in this task,which are difficult to model long dependencies of features and retain global contexts information,the theoretical principle of non-local operation is analyzed,and NAL-Net,a fundus image classification network based on nonlocal attention learning,is proposed.The effectiveness of NAL-Net and the effect of introducing global features through the selfattention learning on network performance are demonstrated by comparing the advantages and disadvantages of traditional algorithms with experimental results on three publicly available fundus data sets.Research on fundus image segmentation method based on hierarchical self-attention.For the problems that U-Net inevitably dilutes global contexts information in the prediction processing and the feature fusion between different levels is simple and rough,hierarchical self-attention module and feature fusion module are proposed.Then the segmentation network for fundus images,HSU-Net,is designed based on U-Net,and the segmentation targets include optic cup,optic disc,blood vessels and lesions.Finally,experimental results on several fundus image datasets with different tasks show that HSUNet performs better overall compared with other segmentation models and can obtain more accurate segmentation accuracy.Research on multi-task learning for retinopathy-assisted diagnosis.Based on the above research,two closely related tasks,DR grading and lesion segmentation,were simultaneously performed in a single model by designing a semi-supervised training method and a multi-task learning network.First,pseudo-labeled data and training sets for semi-supervised training were obtained by using previous research results.Then loss functions and multi-task learning networks are designed to output DR grading and lesion segmentation results in an end-to-end way using output layers at different position.Finally,the proposed method in this chapter is proved to be effective by comparing the experiments with existing mainstream network models.