A Diabetic Retinopathy Classification Method Based On Novel Attention Mechanism

Diabetic retinopathy is one of the most serious and common complications of diabetes.Currently,expert diagnosis relies on the analysis of focal points on retinal images to give diagnostic results,but due to the low manual efficiency,computer-aided diagnostic techniques can help ophthalmologists to achieve diagnosis quickly and effectively.Therefore,this thesis conducts a study on the classification of diabetic retinopathy based on computer-aided diagnostic techniques.In order to solve the problems of high resolution,color interference and small number of lesion samples in retinal images,normalization and different data augmentation methods are used to preprocess the image.In this way,redundant information in the image is removed,the number of samples is expanded and the diversity of samples is enhanced,so as to improve the generalization of the model.In order to solve the problems of sparse focal points in retinal images,and large intra-class differences and small inter-class differences in datasets,this thesis proposes a deep neural network feature extraction method based on a novel pixel-level attention mechanism.The backbone networks are used as the retinal image feature extractors where the pixel-level attention modules are introduced respectively to process the strength of each pixel on the image,thereby paying more attention to important detail features.And through feature fusion,complementary high-level semantic feature descriptors are formed to increase effective information expression,so as to improve the accuracy of the two-classification model.There are some distinguishability characteristics between normal images and diseased images,but the difference is small between classes in diseased images.To solve this problem,this thesis proposes a two-level classification system.The first-level classification model is used to predicted whether the retinal image is normal or not.A multi-task deep neural network structure is proposed in the second-level classification model for learning feature dependencies among different categories to further improve the performance of the multi-classification model.In this thesis,firstly,a novel pixel-level attention mechanism is proposed for the difficulties of the two-classification task,and have better results when compared with different methods on Eye PACS and Messidor datasets.Secondly,a two-level classification system is proposed for the difficulties of the multi-classification task,and the experimental results on Eye PACS and APTOS datasets show that the proposed method can effectively improve the classification accuracy and has good application value for the clinical diagnosis of diabetic retinopathy.