Research On Grading Algorithm For Diabetic Retinopathy Based On Multi-scale Attention Mechanism

Diabetic Retinopathy(DR)is one of the most common ophthalmic complications in diabetic patients and one of the main factors that cause severe impairment of human vision,which has a serious impact on patients’ daily lives.At present,the medical community mainly relies on ophthalmologists to interpret color fundus images to diagnose DR and its severity.This approach takes a long time and relies too much on the doctor’s experience,so it’s not uncommon for missed or misdiagnosed diagnoses.In addition,in some areas of China,due to limited medical resources and levels,many DR patients cannot get timely diagnosis and treatment.Therefore,computer-aided DR intelligent diagnosis and grading technology is of great significance for assisting doctors in the diagnosis and treatment of diabetic retinopathy.This thesis aims to study the DR automatic grading algorithm based on color fundus images from the two dimensions of image level and lesion level,and the main research content is as follows:Research on DR grading algorithm based on image level:(1)Based on the characteristics of different sizes of diabetic retinopathy,a multilayer feature adaptive fusion network structure is designed,making full use of the different receptive fields of different convolutional neural network layers,so that the network model can adaptively represent disease information of different sizes.(2)Design a multi-scale attention mechanism module based on attention mechanism and multi-scale fusion to enhance the extraction and fusion of key lesion area information in the image,and reduce the interference of useless background information and noise.(3)Design a DR grading method combining cross-entropy loss function(CE Loss)and mean squared error loss function(MSE Loss).By making full use of the respective characteristics and advantages of CE Loss and MSE Loss,the network model can reduce the gap between the wrong grading sample and its true lesion grade while retaining the high grading accuracy.Research on DR grading algorithm based on lesion level:The type and number of lesions on color fundus images are important evaluation criteria for DR grading.In this thesis,a dual-branch DR lesion detection and grading algorithm is proposed,and the lesion detection branch detects fine-grained lesion information such as type,number and location of lesions.The attention mechanism guided by the lesion heat map is designed to integrate the fine-grained lesion location information with the features of the classification network,strengthen the feature extraction of key lesion areas,improve the final grading performance of the model,and provide interpretability for the diagnostic grading results of the model.