Research On The Deep Learning-based Conjunctival Vascular Image Segmentation Method

Diabetes induced diabetes retinopathy(DR)is a common ophthalmic disease and the main cause of blindness in diabetes patients.At present,the diagnosis method of DR is to analyze the retinal fundus image,and many researchers have made in-depth research on the segmentation of fundus image.The method of obtaining fundus image is relatively complex,and the conjunctival image is relatively easy to obtain.And research shows that the vascular morphology of the bulbar conjunctiva of the human eye is also significantly related to diabetes retinopathy.Using conjunctival images for disease analysis is a more convenient way than fundus images.The work of this paper also takes this as the background to study how to segment blood vessels on conjunctival images.This research mainly includes the following two aspects:(1)This paper constructs the conjunctival image data set,and designs a conjunctival image enhancement method for the original conjunctival image.Due to the inconspicuity of vascular morphology caused by factors such as angle,brightness and eye structure,a method for processing conjunctival images is proposed.First,the whole conjunctival data set is aligned through the standardization of size and brightness level,and then the green channel containing the most vascular information is selected through experimental analysis to generate grayscale images.Then,a conjunctival image enhancement method is proposed.The image is enhanced by combining MSR algorithm and CLAHE algorithm,so that the contrast of the image is greatly enhanced without distortion,and the blood vessel as the main part of the image is more prominent.(2)Different from other medical image segmentation problems,conjunctival images have no publicly labeled dataset.To solve the problem of no pixel-level labeling,this paper proposes a conjunctival vascular segmentation network based on domain adaptation.Through unsupervised adversarial learning of conjunctival image and fundus image,the model learned the domain invariant features between domains,and realized the blood vessel segmentation on conjunctival image.Different from the traditional domain adaptation method using the basic convolution neural network structure,this paper takes the U-net network,which is very prominent in the field of medical segmentation,as the baseline network,and analyzes the structure of the U-net network in depth during the experiment,improves the U-net network from the characteristics of the image,injects the CBAM attention module into the network,which is used to increase the weight at the spatial and channel levels,In addition,the residual structure is introduced to solve the problem of training difficulties caused by the deepening of network layers.The experimental results show that the method proposed in this paper has obtained meaningful results on the constructed conjunctival image dataset,and has a certain ability of vascular segmentation.