Research On The Method Of OCT Image Feature Extraction Based On Deep Learning In Classification Of Age-Related Macular Disease

In recent years,with the introduction of the concept of deep learning and the continuous emergence of various computing platforms related to it,it has been widely used in image,speech and natural language processing,and has shown its unique strong advantages,and its application continues to be expanded and extended.In the field of medical image processing and analysis,the application of deep learning technology to computer-aided diagnosis and detection systems has also attracted wide attention from scholars.In this paper,we will use the deep convolutional neural network,a machine model under deep supervised learning,to explore its application value in the classification of age-related macular lesions in the fundus retinal optical coherence tomography scan image.The macula is at the optical center of the retina of the human eye and is the most sensitive part of the vision.Age-related macular degeneration(AMD)usually refers to a degenerative disease of the macula in the retina that is highly positively correlated with age.Age-related macular degeneration has the characteristics of diverse types of lesions,early symptoms are not obvious,late lesions are irreparable,and the incidence is high.It has become one of the major eye diseases of elderly blindness.Therefore,regular inspection and early screenings are prevention or delay of this disease.The key to worsening vision disease.As an important method for screening macular disease fundus retinal imaging,due to the use of optical coherence tomography imaging,a noncontact and non-invasive imaging technology with high resolution,it has become the main imaging method for examining retinal diseases.However,the traditional classification study of age-related macular degeneration usually adopts the classification method with artificially marked features.The implementation of such methods often requires manual participation,which is tedious and time-consuming.The deep learning method can automatically learn image features and establish a nonlinear complex model between input and output.With the help of deep learning technology,we have achieved efficient age-related macular degeneration classification research.The main work and contributions of the thesis include the following aspects:(1)Aiming at the impact of the various types of macular disease,irregular shape of the disease,complex representation and other characteristics on accurate classification,a classification model for macular degeneration named Advanced Residual Network(ARN)based on deep convolutional neural network Res Net50 and combined with focus loss objective optimization function was proposed.The design idea of this model is to make full use of the rich nonlinear structure of the residual network to automatically and effectively extract the high-dimensional and general characterization information of the macular lesion from the optical coherence tomography of the fundus retina.The focus loss function is used as the objective optimization function to solve the problem of unbalanced samples in the data set and difficulty in training difficult samples.The experimental results show that our proposed ARN classification model is significantly better than the traditional classification methods of HOG-SVM and VGG16 in the classification performance of age-related macular degeneration.(2)Because the problem that the classification accuracy is not high enough when classifying complex macular disease images,image features and multi-scale information cannot be fully extracted and utilized,the residual network unit as the basis and the attention mechanism are introduced to construct residual attention macular disease classification model(Atten-Res Net).The design idea of this model is to fully utilize the deep network to extract the features of high-dimensional macular lesions,and combine the spatial attention information and channel attention information of the macular disease image through a hybrid attention mechanism to achieve attention weighting of the entire image,to complete the selection of lesion features,so that the model focuses more on effective classification features,suppresses invalid features,and reduces the impact of complex tissues on classification performance.Experimental results show that after introducing the attention mechanism,the Atten-Res Net classification model can effectively improve the classification accuracy of macular diseases.