Research On The Method Of Capillary Segmentation In Microscopic Image Based On Generative Adversarial Network

In the study of diabetic retinopathy,retinal microvascular disease will occur.And hypertension,kidney disease,leukemia,anemia,pregnancy induced hypertension syndrome can cause typical changes in fundus characteristics,fundus morphological characteristics have become an important basis for the diagnosis of many diseases.Clear blood vessels can be seen through the bulbar conjunctiva.Therefore,the accurate segmentation of human conjunctival microvessels can help doctors diagnose the patient’s condition.In view of this,this paper proposes a microvascular segmentation algorithm based on generating confrontation network.The main work of this paper is as follows:The micro image is obtained,and digital image enhancement techniques such as color space transformation,histogram equalization and gamma correction are used to complete the image preprocessing.So as to obtain more suitable data samples for segmentation,and prepare for subsequent segmentation of microvessels.A microvascular segmentation algorithm Ugan based on generated countermeasure network is proposed.The segmentation network u-net is used as the generator of Ugan algorithm to output the segmentation results.Lnet based on convolutional neural network structure is used as the discriminant network of discriminant.Through the confrontation training of two network model minimax game,the optimized generation model can generate better segmentation results.This paper also uses the binary cross entropy loss function and the objective function of the original Gan as the loss function of Ugan.Finally,the segmentation results of Ugan and u-net are compared.The results show that Ugan has better sensitivity.In order to further improve the accuracy of microvascular microscopic image segmentation algorithm,this paper adds channel attention module and dense hole convolution module to the original u-net network,and proposes SCU net network.In this paper,the division network SCU net is used as the generator of the algorithm Ugan,and the scgan division confrontation network is proposed.Scgan algorithm is tested on the fundus image data set Starr.The experimental results show that the segmentation results of scgan algorithm proposed in this paper are ideal.Compared with the gold standard,scgan algorithm can generate more small blood vessels and solve the problem of under segmentation of some small vessels.