Deep Learning Method On Multi-diseases Classification Of Fundus Retinal Image

In the cases of non-congenital visual disability,fundus diseases are the main cause of visual impairment.Diabetic retinopathy,glaucoma,cataract,age-related macular degeneration and hypertension are usually the blasting fuse of fundus diseases.If it can be found and treated in time,the risk of blindness will be reduced,as well as the growth rate of vision disabled patients will be well controlled nationwide.With the upsurge of deep learning technology,computer technology is widely used in many fields,also gradually practicing in the medical field.However,the application of deep learning in ophthalmic disease screening,mostly for single disease,and it’s still relatively scarce for comprehensive diagnosis of multiple diseases,including available databases and reference methods are difficult to obtain.it’s not enough for a fully scale ophthalmic disease screening.The purpose of routine physical examination should be a comprehensive screening,and the health status of patients should be evaluated comprehensively.Therefore,current research results can’t meet the needs of comprehensive screening of fundus diseases.Thus,we propose a multi-diseases classification method based on deep learning and raise a complete solution of aid diagnosis with fundus image.First of all,we normalize the data,cut off the redundant background,and use histogram equalization method to solve the problem of uneven illumination.Then,we enhance and expand the small number of samples to solve the imbalance problem.Finally,we use Resnet50 as backbone,use multi-task learning to solve multi-label classification,and splice the left and right eye images according to the correlation between the left and right eyes.The accuracy of the model is improved by fusing the left and right eye features;because the disease classification of the fundus image is essentially based on the focus on the image,the attention mechanism is introduced to make the model pay more attention to the focus features,and enhance the learning ability of the network to the effective features.The accuracy of the best single model is 0.7069,Jaccard index is 0.5916,F1 score is 0.9232,AUC is 0.9335.Experiment results show that Resnet50 is good at feature extraction and attention mechanism does work.At last,we emerge two models and the accuracy of model fusion is further improved.In our best result,accuracy of model fusion is 0.7103,Jaccard index is 0.6042,F1-score is 0.9261,AUC is 0.9458,which has some auxiliary diagnostic value.The main purpose of this paper is to design a classification strategy and framework for multi diseases and multi tags,rather than discuss the quality of the backbone network,so we choose a relatively simple,rather than the best performance of the backbone network for experiments.In addition,based on the idea of representation layer state transfer(rest),this experiment built an interactive platform for auxiliary diagnosis.The diagnosis program can be run by simple operation to complete the systematic realization of multi disease classification of fundus retinal image.