Research On Image Classification Algorithm For Myopic Fundus Diseases

According to statistics,in recent years,the number of people with myopia in China is constantly expanding,and the degree of myopia is getting higher and higher.When the degree of myopia reaches 600,it is called high myopia,which is also called pathological myopia clinically.High myopia can lead to a many complications,such as "leopard shaped fundus",and may lead to blindness seriously“ "Early detection,early treatment" is the best way to prevent blindness caused by high myopia fundus diseases.However,due to the large-scale increase of high myopia population,it is difficult for existing ophthalmologists to deal with the increasing screening work of high myopia fundus diseases.Using computer technology in the process of diagnosis and treatment of high myopia fundus disease can effectively relieve this situation.In this thesis,we study the automatic diagnosis model of myopic fundus diseases.Firstly,in view of the coexistence of multiple diseases in the fundus images of high myopia,this thesis proposes a GAT-CNN model from the perspective of multi label image classification and considering the correlation between different disease labels.The model is mainly composed of two modules: fundus image feature extraction and label correlation modeling.Convolutional neural network is mainly used to extract the disease features in the image.Because the coexistence of multiple diseases in the fundus image makes the features are more complex,So in this thesis,the features extracted from convolution layers of convolution neural network are mixed as the features of fundus image.Graph attention network(GAT)models the correlation between different high myopia fundus diseases by considering the influence of other labels on the label when updating the characteristics of the disease label each time.The experimental results show that the GAT-CNN model has achieved good results in every evaluation indexes,and the performance of GAT-CNN model in multi label classification of high myopia fundus diseases is better than several multi label classification models of the control group.Secondly,aiming at the imbalance of the distribution of different disease labels in high myopia fundus disease data,from the perspective of cost sensitive learning this thesis redesigns the automatic diagnosis model of myopia fundus disease.Firstly,based on the idea of problem transformation commonly used in multi label classification algorithm,a dual classification network model is proposed.The model solves the problem by a dual classification network for data sets with unbalanced data distribution and a GAT-CNN for data distribution equilibrium.In the dual classification network,cost sensitive learning is combined to improve the recognition ability of the model for minority classes.Due to the existence of two classification networks in the dual classification network model leads the calculation cost is too high,which makes the network difficult to migrate to medical devices for clinical use.Based on this,this thesis proposes a GAT-CNN model based on adaptive cost matrix.The model combines cost sensitive learning in GAT-CNN model,and adjusts the cost matrix adaptively according to the output layer results in the training process,in order to improve the training efficiency of the model.The simulation results show that the proposed both dual classification network model and GAT-CNN model based on adaptive cost matrix can achieve good classification results on high myopia fundus disease data set with unbalanced data distribution.