Research On Intelligent Grading Algorithm For Diabetic Retinopathy

Diabetic retinopathy(Diabetic Retinopathy,DR)is an important factor that makes diabetic patients have a high blindness rate in recent years,and it is also one of the chronic complications with a high incidence rate in diabetic patients.In actual clinical scenarios,manual DR grading diagnosis mainly relies on the images of retinopathy collected by professional ophthalmologists and the clinical experience of professional ophthalmologists.However,DR grading using manual grading diagnosis method not only has a great dependence on the quality of the collected retinopathy images and the diagnostic experience of professional ophthalmologists,but also has high requirements on professional medical resources.Therefore,there are certain limitations in the grading diagnosis method based on artificial DR.With the gradual success of deep learning in the field of medical images,deep learning-based methods have gradually been applied to auxiliary diagnosis of eye diseases,such as DR grade prediction,retinal vessel segmentation,and glaucoma disease diagnosis.Therefore,this paper is mainly based on the deep learning method to construct a retinopathy image quality classification model and a lesion classification model for the research of DR classification auxiliary diagnosis.The main work content of this paper is as follows:1.The quality of images surrounding retinopathy varies greatly,and the generalization ability of the quality grading model is insufficient.This paper proposes a retinopathy image quality classification model based on sharpness perception minimization and multi-color gamut dual-level fusion,which is used to intelligently screen the quality of retinopathy images.The model first uses the Res Ne St network to jointly learn the multi-color gamut features of the three color gamut spaces of RGB,HSV and LAB,and then uses a two-stage fusion module to fully fuse the output feature information and prediction information of the multi-color gamut network.Finally,the sharpnessaware minimization optimization method is used to optimize the quality classification model of retinopathy images to improve the generalization ability of the quality classification model.2.Due to the difficulty in extracting the subtle lesion features of the retinopathy image,and it is difficult to distinguish some lesion feature areas from the background of the retinopathy image.In this paper,combined with the idea of fine-grained image classification,a feature-adaptive filtering retinopathy grading model is proposed to effectively distinguish the grade of retinopathy.The model first uses the feature extraction network to build a multi-scale filtering branch to extract the features of the retinopathy images step by step,and cascades the adaptive feature filtering blocks after different scale filtering branches to enhance and filter the features of the retinopathy images;then uses the feature complementary.The fusion module performs complementary fusion on the multi-scale significantly enhanced features after feature filtering,and uses the complementary information between the multi-scale significantly enhanced features to enrich the global information of the retinopathy image.Finally,the DR hierarchical model is trained by a composite loss function.Qualitative and quantitative experiments were conducted on the open source data sets Eye Q and IDRi D for the model in this paper.Experimental results show that the proposed quality grading model and lesion grading model are superior to a certain extent.