Research On Accurate Auxiliary Diagnosis Method Of Myopia Based On Machine Learning

With the improvement of people’s quality of life,the prevalence of myopia continues to increase,which seriously affects people’s daily life.At present,the diagnosis of myopia-related diseases,including pathological myopia,and the prediction of refractive errors(optometry)are all done manually by doctors.This undoubtedly brings a huge workload to doctors,which leads to misdiagnosis and missed diagnosis.At the same time,in poor areas in the country,with a shortage of experienced ophthalmologists and insufficient large-scale medical equipment,it will be difficult for patients to be diagnosed and predicted in time,which will lead to deterioration of the disease and eventually blindness.Therefore,it is necessary to propose related automatic diagnosis and prediction methods to achieve efficient and accurate diagnosis and prediction without the need of experienced ophthalmologists and large-scale medical equipment.In addition,blood vessels in the fundus can well reflect the severity of myopia-related diseases,and blood vessel segmentation plays a crucial role in disease diagnosis.Therefore,this paper mainly focuses on the automatic diagnosis of pathological myopia,the automatic prediction of myopic refractive error and the automatic segmentation of blood vessels.The main research contents are as follows:1.Research the realization principles of current automatic diagnosis methods for eye-related diseases at home and abroad,including traditional machine learning methods and deep learning methods.Furthermore,summarize and analyze the current existing automatic diagnosis methods for pathological myopia,prediction methods for myopic refractive errors,and fundus blood vessel segmentation methods.2.Research the automatic diagnosis method of pathological myopia.An automatic disc recognition method is proposed to realize fixed cropping of ultra-wide-angle fundus images.In addition,a convolutional neural network named SRM-SE-Dense Net is proposed,which combines SRM and SE attention mechanism to recalibrate the features extracted by the neural network and improve the diagnostic performance of the network.3.Research the prediction method of myopic refractive error.A neural network block AD-Block(Attention Dense)is proposed,and MP-Net(Myopia Prediction Network)is constructed from it.The block is densely stacked by multiple Res-SE-Blocks combined with residual connection and SE attention.After processing the ultra-wide-angle fundus images through image processing,MP-Net is used to predict myopic refractive error and finally obtain the equivalent spherical value.4.Research the method of retinal fundus vessel segmentation.A convolutional neural network MD-Net(Multi-Scale Dense Network)is proposed.The network uses the proposed residual atrous space pyramid pooling block(Res-ASPP)to extract vascular multi-scale information with improved information flow.In addition,the multi-level features in the encoder are densely connected to the decoder to make full use of the features and minimize information loss.Finally,SE-Block is embedded in the encoder to re-calibrate the features after the dense concatenation layer to improve the segmentation performance of the network.