Research On Medical Image Segmentation Based On Convolution Neural Network

Today,medical image is an important auxiliary tool of clinical diagnosis,and using the computer aided diagnosis technology for automatic analysis of medical image can provide an objective diagnosis,which can reduce the influence of doctors' subjective judgment,and improve the efficiency of the doctors' work.For example,breast cancer and retinopathy gradually become common diseases,and have a great influence on human health.If we can use the computer aided diagnosis technology to early screening for these diseases,it is beneficial for the diagnosis and treatment of disease.As for traditional medical image processing method of image segmentation,researchers firstly conduct image segementation experiment,and then extract the lesion area characteristics.However,there are many different kinds of medical images,which each of has different characteristics,and lesion areas are complicated.Therefore,segmentation has been an important and challenging issue in medical image processing.In recent years,deep learning has attracted more and more researchers' attention due to its stronger learning ability and Convolutional Neural Network(CNN)is a classical deep learning model for processing images.So,in this paper,we mainly conduct a study on breast ultrasound images and fundus images to learn the segementation of medical image based on the deep learning.The main contents are the following:1.This paper proposes a hierarchical framework for breast tumor segmentation,consisting of two models which are the low-level segmentation model and the high-level segmentation model.Firstly,based on the local level gray clustering hypothesis,level set which is a kind of activite contour model is used as the low-level segmentation model for initial segmentation,and the we can find the misclassified regions according to artificial annotation.Then the result of low-level model can be corrected by the high-level segmentation model which is based on convolutional neural network.In this model,we put the misclassified regions form the low-level segementation model into the CNN model,and train the network model to recognize the specific region in the breast image in order to improve the segmentation performance.The high-level segmentation model is learned from the low-level segmentation results,so it has good complementarity and coupling with the underlying segmentation model,and it can improve segmentation accuracy.We conduct experiments on the self-build breast ultrasound database,and it can show the effectiveness of this method.2.In this paper,we propose a method based on the combined convolution neural network to acquire the exudate regions of retinal images.The network consists of pre-segmentation network,hierarchical classification network and exudate regions segmentation network.Firstly,we use the sample to train the CNN network,thus we can get the pre-segmentation network.We compare the segmentation results of the training samples with manual annotation to classify samples to simple grade class label which are correct segmentation region and difficult level class label which are error segmentation region,so we can obtain the prior information from training samples.Then training the grade classification network based on the samples annotated simple or difficult level.The network has the ability to divide grades.Finally,a network for hard-to-segment samples was trained to get the exudate region of the more difficult segmentation samples.In this paper,the proposed combined convolution neural network can firstly find the samples that are not easy to segment,and then use the special network to test the samples to improve the segmentation accuracy of the difficult samples.The experiment is carried out on the open database e-ophtha EX,and the effectiveness of the method is proved by the experimental results.