Research On Auxiliary Diagnosis Technology For Pathological Myopia Based On Deep Learning

Pathological myopia is an extremely serious form of myopia,the incidence of which has been increasing in recent years,and the age of the population at risk is gradually decreasing,posing a major threat to human eye health.The traditional clinical diagnosis of pathological myopia requires a doctor’s personal experience and expertise to make the diagnosis.However,this approach is not only limited in its efficiency,but also carries the risk of subjective error.At the same time,in areas where there is a shortage of ophthalmologists,there is a high risk that patients will not receive timely diagnosis and treatment,which can have a significant impact on their recovery.With the development of the intersection of computer and medical disciplines,deep learning-based assistive diagnostic techniques have demonstrated great potential.In the diagnosis of pathological myopia,the fundus image is the most crucial basis,in which the features of lesions appearing in the fundus structures such as the optic disc,retina and choroid are clearly presented.At the same time,the blood vessels in the fundus are also important in the diagnosis,as their morphological information can show the nature and extent of the lesion and help the doctor to make diagnostic decisions.Accordingly,this paper presents research work on fundus image processing,fundus image vessel segmentation,and pathological myopia fundus image recognition,using fundus images as the data base,with the following main work:(1)The method of fundus image processing.A fundus image processing method that combines contrast-constrained adaptive histogram equalisation(CLAHE)with a denoising neural network(Dn CNN)is proposed to address the problem of variable fundus image quality and the lack of dedicated denoising during traditional fundus image pre-processing.It reduces the noise level while enhancing the image contrast and making the fundus structures more visible.In particular,the number of network layers,the activation function and the edge filling method of the Dn CNN model were modified to suit the fundus image denoising task in this paper.(2)The method of fundus image vessel segmentation.An improved U-net-based vessel segmentation method is proposed to address the problem of inaccurate segmentation of fine vessels and vessel endings that occurs in existing segmentation methods.The improved U-net consists of the Dconv Block structural block with deformable convolution and channel attention introduced,and the Block structural block,which can pay more attention to vessel information when performing feature extraction in the downsampling stage.Also,the batch normalisation layer was used in Dconv Block and Block to improve the stability of the model.In addition,the change to Soft Pool as the pooling method of the model reduces the information loss during downsampling.(3)The method of pathological myopia fundus image recognition.A recognition method based on the improved Efficient Net V2-S model is proposed to address the problem that the overall accuracy of existing recognition methods is generally low,especially for positive samples.The improved model has a more balanced number of convolutional kernels and network layers,and the SE attention is replaced by ECA attention,which makes the calculation of attention weights more accurate.At the same time,the Adam optimiser is used as an optimisation method to avoid the model falling into local optima and to allow the model to converge more quickly.In summary,this paper investigates aided diagnosis techniques for pathological myopia based on deep learning,and proposes three improved approaches for fundus image processing,fundus image vascular segmentation,and pathological myopia fundus image recognition,respectively,to address existing problems.Through in-depth analysis of fundus images,these methods can effectively improve the diagnostic efficiency and accuracy of pathological myopia.