| Diabetic Retinopathy(DR)is the leading cause of blindness in adults.As the population of diabetes continues to increase,the number of patients with diabetic retinopathy also increases year by year early.The early detection of this condition is essential for good treatment.The retinal fundus image is an important basis for doctors to diagnose eye diseases,and accurate analysis of the pictures can enable patients to be treated in time.However,traditional manual diagnosis methods are time-consuming and labor-intensive,and at the same time limited by the lack of medical resources,large-scale rapid diagnostic screening cannot be effectively performed,which ultimately leads to many patients’ visual impairment due to untimely treatment.This places a huge burden on individuals and society.In recent years,convolutional neural network(CNN)has been widely used in the field of image recognition.Using the model of convolutional neural networks can help ophthalmologists make quick and effective diagnosis.This can save medical resources and allow patients to be treated in a timely manner.This is of great significance for the prevention and treatment of fundus diseases.This paper hopes to carry out multi-feature learning on fundus image data by using the method of deep learning to effectively classify and diagnose diabetic retinopathy.Accordingly,this paper describes the integrated classification model of diabetic retinopathy based on fundus image,which mainly includes the following two parts:First,this paper proposes a fundus images of image information processing algorithm,Rotate Cut Mix.In this paper,through targeted pre-processing of two fundus data sets from different sizes and data sources,the feature information of the original medical data sets with small sample sizes and large differences in image quality was enhanced.Meanwhile,the correlation information between different features could also be enhanced.In addition,the Rotate Cut Mix can provide a discriminant basis for subsequent diagnostic classification and complete disease classification through the reuse of multiple features.Second,this paper constructs an integrated classification model for retinal fundus image based on convolutional neural network and this model can use the fundus image data to learn features and further complete the classification of disease course.In this integrated classification model,each submodel is used to screen the required features for correlation,and then the feature information that is most relevant to the disease is obtained and output according to the classification.By modifying the network structure and loss function of each submodel,this model innovatively transforms the fundus image classification problem into a regression problem,realized the combination of convolutional neural network and traditional machine learning method,and improved the recognition rate of the model to 91.36%.In addition,this paper adopts a transfer training method based on machine learning,which reduces the computing resources needed for training by transferring and learning the features,improves the learning speed of the network,enhances the ability of feature expression,and improves the classification performance of the model.In this paper,the features extracted from the convolutional network are analyzed by the hot spot diagram,and the classification basis of the integrated model is explained intuitively.Compared with other methods,this paper USES five-fold cross validation to compare the results.The experimental results show that the model training in this paper costs fewer computing resources and has higher classification accuracy.The accuracy of classification integration model is 91.36% and F1 is 86.42%,which is nearly 5% higher than the previous method.In this paper,innovative image information processing algorithms,different convolutional neural networks and feature selection methods are used to conduct targeted deep learning and classification of color fundus image data.The method proposed in this paper achieves excellent classification performance by using less computational resources. |
