Research And Implementation Of Segmentation Technique On Retinal Fundus Image Based On Deep Learning

In clinical medicine,retinal fundus image is an important basis for ophthalmologists to diagnose and treat patients with fundus diseases.Oph-thalmologists conduct detailed screening of latent diseased areas in fundus images to determine the patient's condition*.Diabetic retinopathy,for ex-ample,ophthalmologists can diagnose the patient's condition by analyzing the structure of the retinal fundus image.The hard exudate is one of the early manifestations of diabetic retinopathy,and the shape of the retinal blood vessels can reflect the degree of the diabetic retinopathy.This thesis studies the segmentation algorithm of retinal fundus image based on deep learning technology,and designs neural network models based on deep learning to assist medical diagnosis.The main contents are as follows:(1)Investigate and sort out the retinal fundus image dataset.For the different characteristics between hard exudate and retinal blood vessels,this thesis adopts digital image processing technologies,for example,color space conversion,histogram equalization,image enhancement and so on in order to obtain better training data samples and prepare for subsequent model training.(2)Design a hard exudate segmentation model based on multi-scale feature fusions.This model adopts encoding-decoding network structure,extracts more information through multi-scale feature extraction module,and adopts residual network module to make the network structure transfer information well when it grows in depth.False positive bootstrapping strat-egy is used to optimize the training process and increase the accuracy of the network model.The hard exudate model has a certain degree of im-provement in sensitivity and performs segmentation tasks well,which can be used to assist medical diagnosis.(3)Design the retinal vascular segmentation model based on the gen-erative adversarial network.The generator model adopts encoding-decod-ing network structure,the output is the result of retinal vascular segmenta-tion.The discriminator model has classic structure of convolutional neural network,the output result is the classification of the input image,that is,to judge whether the input image belongs to ground truth or the generator model segmentation result.By means of the alternate iteration of two net-work models,the segmentation results of the generator model are opti-mized.The final model is tested on DRIVE and STARE dataset,and the UNet network model,which has been initially applied in the field of retinal vascular segmentation is used for comparison.The results show that the model based on generative adversarial network has a certain degree of im-provement in sensitivity,and the segmentation results have a clearer per-formance in details.