Research On Classification Of Retinal Fundus Diseases Based On Deep Learning

Retinal fundus disease is not only diverse,but also have great damage to visual function.Common retinal diseases include age-related Macular Degeneration(AMD)and diabetic Macular Edema(DME).These diseases can damage the retina and cause vision loss,and are the most common cause of blindness in the elderly,so its early detection is of great significance to follow-up treatment.At present,clinical diagnosis of the above two retinal diseases depends on professional ophthalmologists.However,with the increasing number of patients,it will be a very difficult task to rely on only a few professional ophthalmologists for such a large number of diagnosis and screening work.Based on the above problems,this paper adopts the methods of machine learning and deep learning to classify and diagnose retinal fundus diseases with the help of computer vision and other technologies.Most of the existing diagnoses of retinal fundus diseases are based on images obtained by fundus photography and small-field Optical Coherence tomography(OCT).Compared with fundus photography,OCT images are characterized by non-invasive,high speed and high resolution.This study is based on the classification and recognition of retinal fundus OCT images.The specific research work is as follows:(1)an OCT image classification method based on classic machine learning is studied,which mainly uses the Support Vector Machine(SVM)algorithm and K-Nearest Neighbor(KNN)algorithm to identify the retinal fundus disease images,respectively.The experimental data comes from the OCT dataset(including AMD,DME and NORMAL images)provided by Duke University,and using the ideas of cross validation to authenticate the data on the test set.(2)An OCT image classification method based on 3D Convolutional Neural Network(3D-CNN)model is proposed.The differences between OCT images are not obvious,and there is a lot of speckle noise,which makes it difficult for traditional image classification algorithm to achieve a more accurate classification effect.According to the strong learning ability of the convolutional neural network,the improved 3D-CNN model OCT image classification algorithm adds the Inception module,and achieved a good classification effect on the model evaluation index.(3)In view of the basis of convolutional neural network,an OCT image classification method based on transfer learning was proposed for the small data set of retinal fundus disease OCT images.The pre-trained deep learning framework VGG-16 was used to this time,fine-tuning its deep network parameters and reconstructing the fully connected layer,and migrate it to the retinal disease OCT image classification task.The above experimental results show that the two deep learning methods based on convolutional neural network have better classification results than the machine learning methods.Among them,the OCT image classification method based on transfer learning can better distinguish the OCT image of retinal fundus diseases from the normal image and achieve the best classification accuracy.