The Study Of Medical Image Segmentation Based On Convolution Neural Networks

Image segmentation is one of the key technologies in medical image processing,and it is an important reference for doctors to diagnose diseases and make treatment plans.For example,doctors may judge whether a patient has glaucoma based on the shape,color and luster of the optic disc in the fundus image,or whether a tumor is benign or malignant based on the location,size and shape of the brain tumor.At present,various methods can automatically segment medical images,but convolutional neural network has gradually become the most mainstream segmentation method due to its end-to-end processing and robustness in various medical modal data and organ segmentation.Especially in recent years,with the rapid development of GPU processing power,the data processing of convolutional neural network can be realized in real time,which increases its application in the field of medical image segmentation.However,the existing methods of convolutional neural network for image segmentation still face many problems,such as some can only focus on the whole or part,but not combining them,and it's easy to miss the small areas.In view of the shortcomings of medical image segmentation based on convolutional neural network,two kinds of automatic segmentation models are proposed to realize the automatic segmentation of fundus image and brain tumor MR image.The main contents and innovations of this paper are as follows:1.Aiming at the problem of optic disc segmentation in fundus image,the Generative Adversarial Nets(GAN)is used to realize the accurate segmentation of optic disc.In this study,the Unet framework is used to build a generator.It can combine the high-level semantic information and low-level semantic information in the image to segment the fundus image,which solves the problem that the image segmentation accuracy is not high due to the complexity of the grayscale distribution and boundary distribution of fundus image.The discriminator is designed with full connection layer.The input of the network is not simply to output the result of the generator.The input of the discriminatory network is composed by two part.The first part is the result of that we use the output of generative network as a mask to filter the original fungus image.The second part is the result of that we use the ground truth as the mask to filter the original fungus image.The output of the network is no longer simply to determine the true or false of images,but to calculate the difference between the two images.Through such operation,the training convergence of the Generative Adversarial Nets is better,and the precision of segmentation is also higher.2.To solve the problem of brain tumor MR image segmentation,the Edge-Net model is proposed.The model can not only extract the global information in the image but also accurately segment the boundary of the key region.Three subnetworks of edge-net are constructed,which are edge subnetwork,attention subnetwork and segmentation subnetwork.The edge subnetwork can locate the tumor region and determine the tumor boundary.The attention subnetwork transforms the tumor boundary information into the weight information and increases the weight of the convolution network at the tumor boundary.At the same time,the attention module is used to divide fine organizations precisely.The global feature information of the whole image is extracted by the dilated convolution.The image feature information extracted from the attention subnetwork and segmentation subnetwork is spliced together to predict the segmentation result.On this basis,this paper applies the ResNet model to extract all feature information from the input image,and further improves the ability of the ResNet model to extract features by using the method of combining the high level semantic feature and low level semantic feature,thus improving the accuracy of network segmentation.