Blood Vessel Segmentation And Detection Of Pathological Myopia Based On Deep Learning In Fundus Images

In recent years,with the continuous improvement of the living conditions of the whole society,the popularization of electronic equipment and the improper use of the eyes in daily life,the proportion of people with myopia continues to increase,especially the prevalence of myopia among teenagers is generally high,which has become a worldwide problem of eye health.If attention is not paid to the increasing degree of myopia,mild myopia may continue to worsen into pathological myopia with a high rate of blindness.At present,visual acuity examination is time-consuming,laborious and requires professional doctors,which brings great challenges to the screening and diagnosis of eye diseases.Therefore,the realization of automatic fundus image segmentation and pathological detection has important application value for clinical medicine.In this thesis,based on the research hotspot and difficulty of retinal fundus images,retinal blood vessel segmentation,optic disc segmentation and pathological myopia detection in fundus images are realized based on the deep learning framework.The main research contents are as follows:(1)In order to solve the problem of fracture of small blood vessels during segmentation,this thesis studies and implements a retinal blood vessel segmentation method based on improved U-Net multi-scale fusion.First of all,this thesis obtained the latest fundus image data from the clinic,completed the self-made data set Dataset100,and expanded the data set of retinal blood vessels.Secondly,methods such as gray transform,normalization processing,CLAHE enhancement and Gamma correction were used to preprocess retinal blood vessel images.In addition,in order to prevent data overfitting caused by a small number of samples,this thesis also adopts the patch method for training.In addition,the retinal vascular segmentation network based on improved U-Net multiscale fusion is based on the original U-Net network,introducing Coordinate Attention module,Res2 Net Block module and cascade cavity convolution.Experiments on DRIVE,CHASEDB1 and Dataset100 show that the accuracy are 96.90%,97.83% and 94.24%,respectively.The AUC values were 98.84%,98.98% and 97.41%,respectively.Compared with U-Net and other mainstream methods,the sensitivity,accuracy and other indicators were improved,indicating that the vascular segmentation method in this thesis has the ability to capture complex features and has higher advantages.(2)Optic disc segmentation is of great significance in the detection of fundus lesions.In order to solve the interference caused by the existence of blood vessels and other conditions to the optic disc and realize the accurate segmentation of the optic disc,this thesis studies and implements a multi-scale input based residual U-Net optic disc segmentation method.Firstly,the brightness of the optic disc region was enhanced by CLAHE enhancement and Gamma correction.Secondly,the multi-scale input based residual U-Net optic disk segmentation network takes U-Net as the backbone network,adding residual structure,pyramid pool module and multi-scale input module into the network.Finally,experiments on the ORIGA open data set of fundus optic disc show that the proposed method not only eliminates the interference of blood vessels,but also has good segmentation performance,where the overlap error is 0.105 and the balance accuracy is 0.976.(3)Research and implement a pathologic myopia detection method based on deep learning.First of all,in order to eliminate the accuracy of pathological myopia detection due to the interference of dark spots and shadows in fundus image data,this thesis used brightness transformation,contrast enhancement and other methods to preprocess the data.Secondly,since the optic disc region contains rich information features of pathological myopia,this thesis clipped the optic disc region of the picture by using the PPYOLO Tiny model,and put the clipped picture into the model for training.Finally,the pathological myopia detection framework proposed in this thesis mainly consists of three parts,that is,the mature image processing network Res Net50 is taken as the backbone network,and ECA attention mechanism module and grouping convolution module are introduced into its structure.The accuracy,specificity and sensitivity were 79.27%,92.78% and 68.08%respectively on PALM data set.