Diabetic Retinopathy Detection Based On Multiple Transfer Learning

Diabetic retinopathy,one of the serious complications of diabetes,has aroused people's attention due to its high incidence and blindness rate.In traditional research,the collected fundus images of patients are mainly classified by machine learning,and the clinical experience of doctors is used as auxiliary judgment.However,due to the uneven technical level of doctors and low accuracy of detection by machine learning,the vast majority of patients do not receive regular screening or fail to receive effective treatment at the early stage of the disease,resulting in visual impairment and even blindness.Deep learning and transfer learning have made remarkable achievements in the field of image processing in recent years.Deep learning can extract features from the data to improve the detection accuracy of the model,while in the training of deeper convolutional neural networks,it is still dominated by large-scale labeled data sets.Based on fundus images and taking diabetic retinopathy as the research object,this paper establishes an algorithm model to classify the lesion grade.The main contributions of this paper are as follows:(1)A series of fundus image preprocessing and image enhancement methods are proposed.During the construction of the initial data set,a variety of image preprocessing methods are proposed to solve the noise problem of fundus images of patients,avoiding the influence of various subjective and objective reasons on lesion detection during image acquisition.When constructing the secondary data set,various unbalanced methods of image expansion and enhancement methods of lesion feature are adopted.Experiments show that the accuracy of the model can be improved by preprocessing images in various ways.(2)A method of introducing pre-training model and multiple transfer is proposed.The pre-training models VGG19,Inception-V3 and Res Net50 trained on Image Net are introduced to call their shallow network weights.Experiments show that the performance of the three models has been improved to a certain extent after the introduction of the pre-training model.The pre-training models are trained by using the constructed secondary migration data set,and their performance is improved compared with the primary migration learning,which shows that richer levels can also improve the model performance to a certain extent.(3)Several optimization methods are proposed.A detection method for diabetic retinopathy based on integrated learning is proposed and different network models are integrated to improve the model performance.A method lacking function weighting is introduced to solve the problem that the number of negative samples is too small due to unbalanced data distribution in the data set.Inception-V3 model was designed to train fundus images with different resolutions,and to explore the influence of different resolutions on model accuracy...