Study On Classification And Evaluation Model Of Diabetic Retinopathy

Diabetic retinopathy is a possible complication of late diabetic patients,and has become the main cause of impaired vision and blindness in diabetic patients.The realization of automatic diagnosis and classification of diabetic retinopathy is beneficial to the prevention of visual impairment.Manual diagnosis and classification rely on the observation and statistics of various lesions（microaneurysm,exudation,hemorrhage,etc.）in color fundus images by the unaided eyes.However,due to the complexity of retinal structure and weakness of lesions,manual diagnosis is a heavy task with low accuracy.Therefore,it is of great challenge and practical significance to realize the automatic assessment of retinal disease and assist doctors in medical diagnosis through computer vision technology.Based on color fundus images,the algorithms for diagnosis of diabetic retinopathy and automatic classification of lesion grade were designed in this thesis.The main research contents are as follows:（1）A method for detecting microaneurysm based on multi-feature combination was proposed in this thesis,which realized the detection of the initial symptoms of diabetic retinas.Target salience features and texture features can only describe the object itself,but it is difficult to fully describe the local structure characteristics of its background,this thesis proposed the ring gradient descriptor（RGD）to measure the difference of the target and the background,and then all features are combined and fed into gradient boosting decision tree for classification.Finally,the model was tested on the ROC dataset and the Ephtha-MA dataset.The RGD algorithm improved the detection performance of microaneurysm.The AUC value was improved from 0.9066 to 0.9409 in the ROC dataset,and the AUC value was improved from 0.9615 to 0.9752 in the Ephtha-MA dataset.The proposed algorithm finally achieved FAUC values of 0.356 and 0.630 in two datasets,respectively,and the test results in the Ephtha-MA dataset reached the leading level.（2）The application of deep learning in color fundus images was studied.The classification performance of convolutional neural network Res Net34,Res Net50,and Dense Net121 under various parameter settings was tested in the open dataset APTOS2019.（3）A model of diabetic retinal grading based on fine-grained classification was proposed.Traditional fundus image classification model based on deep learning only relies on depth features of image level,but key lesions of different levels only occupy a small part of pixels in the image,resulting in poor interpretability and poor generalization ability of the model.In this thesis,candidate areas of lesions were extracted from the original image according to the characteristics of various lesions and reassembled into new lesions,which were used as the input of the model together with the original image.The image features and fine-grained lesion features of various lesions reassembled images were extracted and used for classification through specific structural design.The proposed method was tested in the public dataset APTOS2019 and the actual dataset,and the model performance was improved after the introduction of fine-grained features.In the open dataset APTOS2019,Kappa value increased from 0.8947 to 0.9164,while in the actual dataset,Kappa value increased from 0.6719 to 0.8352.