Research On Intelligent Aided Diagnosis Model Of Common Ophthalmic Diseases For Color Fundus Images

Nowadays,the incidence of fundus diseases in China is increasing year by year.Due to the specialization of ophthalmology and the lack of professional ophthalmologist resources,the coverage of primary hospital treatment is extremely limited,and it is difficult to meet the current demand for treatment of fundus diseases by only increasing the training of ophthalmologists.In recent years,deep learning technology has flourished in the field of fundus image recognition,and it has become a trend to study deep learning-based ophthalmic disease recognition models that can help doctors make preliminary diagnosis of ophthalmic diseases.At present,deep learning-based ophthalmic disease recognition models mainly focus on single ophthalmic disease recognition,and there is less research on multiple ophthalmic disease recognition models.At the same time,deep learning models require a large amount of existing labeled data for model training,and collecting and labeling ophthalmic disease data is a very arduous task.In this study,we construct a multi-classification recognition model for common ophthalmic diseases based on deep learning techniques,and perform data augmentation on existing ophthalmic datasets by class activation localization methods,and finally introduce an attention mechanism to further improve the recognition performance of the multi-classification recognition model.The main research contents are as follows.（1）Based on migration learning,six classification ophthalmic common disease recognition models based on VGGNet16,Inception V3,Res Net50,Dense Net121 and Efficient Net B0 were trained,and then the results on the recognition models obtained by training on the ophthalmic common disease dataset before and after preprocessing were compared.The experimental results show that among these five models,the Efficient Net B0 model based on migration learning has the best effect in recognizing common ophthalmic diseases,and the recognition effect of each model obtained by training on the pre-processed dataset has been improved to some extent.（2）To address the problem that the data set was not significantly enhanced by the basic data enhancement method,deep learning interpretability techniques were introduced to visualize the ophthalmic common disease recognition models by class activation,and then the ophthalmic common disease data set was enhanced by the class activation localization method.The experimental results show that the diagnostic effects of the ophthalmology common disease recognition models after data augmentation by class activation localization method are all improved,among which the improvement effect is significant for Inception V3 and Res Net50 models.（3）To address the problem that the accuracy of the ophthalmology common disease model after data enhancement is still low,the Inception model and Res Net model with obvious improvement effect after data enhancement are fused,and the Res Ne Xt model based on Res Net with Inception as the grouping architecture is introduced to improve the ophthalmology common disease recognition model by the attention mechanism.The experimental results show that the recognition model of common ophthalmic diseases fused with Res Ne Xt and attention mechanism has higher accuracy in recognizing common ophthalmic diseases than the previous Efficient Net B0 model with the best results.This study provides technical support for the subsequent deployment of the model in primary hospitals,which helps the primary hospitals in the initial diagnosis,accurate classification and timely referral of common ophthalmic diseases.