Research On The Diagnosis Of Fundus Lesions Based On Multi-label Classification

Fundus disease is a serious eye disease,including glaucoma,cataract,retinopathy,etc.If it is not treated in time and effectively,it will cause irreversible damage to vision and even cause blindness.Therefore,early and accurate diagnosis of fundus diseases is the key to avoid visual impairment and treatment.However,fundus diseases are atypical and difficult to detect in the early stage,and manual diagnosis has the disadvantages of time-consuming,laborious,and strong subjectivity.With the development of medical big data and artificial intelligence technology,it is possible to automatically diagnose fundus lesions.However,the existing automatic diagnosis algorithms for fundus lesions have shortcomings such as low accuracy,high missed diagnosis rate,and ignoring the correlation between lesions.The early lesions of fundus diseases are small,and there is a certain correlation between different lesions,which poses a high challenge for the automatic diagnosis of fundus lesions with high accuracy.This article focuses on the automatic diagnosis of three common fundus lesions: glaucomatous optic disc changes,retinal exudates and hemorrhages.Aiming at the problem that the characteristics of fundus lesions caused by early fundus diseases are small and mostly distributed at the edge of the retina,this paper proposes an automatic diagnosis method for multiple fundus lesions that integrates a dual attention mechanism.The method includes three parts: data preprocessing,feature extraction and fundus multiple lesion diagnosis.Data preprocessing first uses the Hough transform algorithm to extract the circular boundary of the retina,automatically crops out the smallest circumscribed rectangle containing the circle to remove surrounding noise,and then uses image normalization,data enhancement and other technologies to improve the diversity and quality of data,thereby enhancing the robustness of the network.The convolutional neural network(CNN)used in feature extraction adopts a residual structure with skip connections.Grouped convolution is utilized in the residual blocks to reduce the number of network parameters,and channel and spatial attention mechanisms are embedded after each group convolution to improve the accuracy of fundus disease diagnosis.The diagnosis of multiple fundus lesions is achieved by training a diagnostic model for each lesion and integrating them together.The method was tested on the clinical data of Ningbo Eye Hospital Affiliated to Wenzhou Medical University.The diagnostic accuracy rates of glaucomatous optic disc changes,retinal exudation and hemorrhage were0.982,0.975 and 0.972,respectively.The experimental results show that the method has better feature extraction ability and diagnostic performance in the diagnosis of fundus multiple lesions.Aiming at the clinical problems that one fundus disease often causes multiple fundus lesions,there is a certain correlation between different fundus lesions and the difficulty of modeling lesion correlations,this paper proposes an automatic diagnosis method for multiple fundus lesions based on a feature sequence processing model Res Net-LSTM and DGNNDS,an automatic diagnosis method for fundus multiple lesions based on deep graph neural network(GNN).Res Net-LSTM first uses Res Net to extract high-level semantic features of fundus images,which are then passed to Long Short-Term Memory(LSTM)networks to explore the correlation between different lesion feature sequences.Finally,the output features of Res Net and LSTM are fused,and a multi-label classifier is used to achieve automatic diagnosis of multiple fundus lesions.DGNNDS first uses a deep CNN to extract high-level semantic features of fundus images,which are then input to a GNN to establish correlation models between different fundus lesions by mining and learning their correlations.Finally,the output features of the CNN and GNN are fused,and a multi-label classifier is used to achieve automatic diagnosis of fundus lesions.The two methods proposed in this paper can improve the diagnostic performance of the model by learning the correlation between fundus lesions,and are superior to advanced Res Net101 and Dense Net121 algorithms in both qualitative and quantitative evaluations.The experimental results tested the proposed models and compared them with other algorithms in terms of average F1 score per class,average overall F1 score,average accuracy of all categories,and the area under the receiver operating characteristic curve.Specifically,Res Net-LSTM achieved 0.911,0.910,0.968,and 0.977,while Res Net-GNN achieved 0.905,0.906,0.959,and 0.974.