Research On Fundus Disease Diagnosis Algorithm Based On Deep Learning

The development of computer and artificial intelligence technologies has played a key role in the progress of human beings.In recent years,the gradual maturation of deep learning technology has made it the first to start landing applications in some fields,such as speech recognition,translation,autonomous driving,abnormal diagnosis and so on.Among them,the application of machine vision in the classification of fundus medical diseases has become a recent research hotspot,becoming a beneficial assistant to assist clinicians in diagnosis.However,due to the problems of complex fundus image features,irregular diagnostic regions of multiple diseases,and extremely high pixel size of the images captured by fundus cameras,the prediction accuracy of models obtained by direct training using current mainstream classification networks is low and the prediction time is also long.To address the above problems,this paper improves the deep learning network in three aspects,such as improving the fine-grained feature extraction ability,massively reducing the number of parameters and computation,and applying large-sized original images to the training network,with the main research contents as follows.(1)To address the problem of complex fundus image features and strong interregional correlations that lead to the low accuracy of current mainstream deep neural networks,research based on the combination of improved Transformer and convolution is carried out.To this end,a Trans Eye fine-grained classification network based on a weighted attention mechanism is proposed.trans Eye combines the advantages of convolutional and Transformer models,which can both effectively extract the underlying features and establish the image remote dependencies,so that the most discriminative image regions can be located and trained end-to-end.The method is experimentally validated on the pre-processed OIA dataset.(2)A study of bilinear involutional neural network architecture based on the involution operator was carried out to address the problem of slow training and prediction due to the large number of computations and parameters in the Transformer structure.To this end,an ABINN classification network is proposed,which has 11% of the number of parameters of a conventional bilinear convolutional neural network(BCNN),extracts the underlying semantic and spatial structure information of fundus images and performs second-order feature fusion,and is an effective parallel of convolutional and attentional methods.In addition,two instantiation methods Attention Subnetwork based on Patch(AST)and Attention Subnetwork based on Pi Xel(ASX)are proposed to implement attention computation based on involution operator,which can accomplish attention within the underlying structure of convolution computation,enabling bilinear subnetworks to be trained and feature fusion performed under the same architecture.Experiments on the publicly available fundus image dataset OIAODIR show that ABINN has an accuracy of 85%,which is 22.5% better than the generic B convolutional model and 0.9% better than the Trans Eye model.(3)Limited by the size of the original fundus medical images,traditional deep neural networks usually compress the original images isometrically before feeding them into the model training,a process that inevitably loses a large amount of detailed information and thus reduces the model accuracy.To solve this problem,this paper conducts a new model study based on multiple example learning and proposes the DFMINN model,which is an end-to-end dual-stream network using the resnet-based multi-instance spatial attention(MISA)module to extract local information and the Global priors network base on involution module to analyze the overall content,respectively.In addition,we use the Double Flow Feature Fusion module to fuse the two information flows and use the ASL loss as the classification head to obtain the prediction results.Experiments on the publicly available multi-label fundus dataset OIA-ODIR show that DF-MINN outperforms previous networks in the prediction of all seven diseases.The ablation experiments further demonstrate the importance of high-resolution images for fundus disease diagnosis tasks.The above three improvement methods increase the accuracy of deep learning visual models in fundus disease diagnosis tasks and save the time and space costs of training and detection,thus having practical application value.