Fundus Retinopathy Image Classification Based On Deep Learning

Early detection of lesions is of great significance for the treatment of fundus diseases.Fundus photography is an effective and convenient screening technique,through which common fundus diseases can be detected.The purpose of this paper is to use color fundus images to distinguish various fundus lesions.The existing fundus disease classification research has achieved some success through deep learning technology,but only using a deep CNN architecture with limited global modeling capabilities,there is still a lot of room for improvement in the evaluation indicators of the classification model,and the simultaneous diagnosis of multiple fundus diseases is still facing challenges.Great challenge.Aiming at the different characteristics of fundus images compared with natural images,this paper focuses on the preprocessing and construction of fundus datasets,the research on feature extractors based on deep convolutional neural networks and Transformer architectures,and the design of backbone networks based on hybrid architectures.research has been carried out.The main research contents include(1)preprocessing of the fundus image data set,including locating the retinal region of the fundus,cropping the black edge of the image;performing morphological transformation on the image,and using histogram equalization for data enhancement to achieve data expansion.It has been verified by experiments that this data augmentation method can effectively improve the fitting degree of the model and significantly improve the classification accuracy.(2)Research feature extractors based on deep convolutional neural networks including Goog Le Net,Res Net,Efficient Net,etc.,and weighted loss functions proposed for class imbalance problems,and verify their respective results on preprocessed fundus image datasets.performance.(3)Study the working principle of the self-attention mechanism and feature extractors based on the Transformer architecture including Vi T and Coa T,and evaluate their performance using the fundus image dataset.(4)Study the hybrid architecture model combining convolutional neural network and Transformer architecture,and propose a feature extraction method combined with Transformer self-attention module for the precision loss of convolutional neural network due to the lack of global receptive field,convolution block extraction The local information of the fundus image is extracted,and the complex relationship between different spatial positions is further captured by the self-attention module,and one or more fundus diseases in the retinal fundus image can be directly detected.In the initial stage of feature extraction,this paper proposes a multi-scale feature fusion stem structure,which uses convolution kernels of different scales to extract low-level features of the input image and fuse them to further improve the recognition accuracy.The experimental results show that,compared with some proposed network models based on a single architecture or a hybrid architecture,the hybrid model we propose achieves state-of-the-art performance in multiple metrics such as F1-score,Kappa coefficient,and AUC value with fewer parameters.