Research On Medical Image Segmentation Based On Neural Network And Active Contour

Medical image segmentation is the essential link of medical diagnosis and image analysis,which provides support for medical organization research and clinical diagnosis.However,medical images are prone to noise and gray level inequality due to the influence of equipment and human internal structure.Moreover,different organ sizes and shapes of different patients are discrepant,which brings challenges to the nidus segmentation of medical images.The existing active contour model method does not need training set and can make good use of the edge information of the target,but it is sensitive to the initial contour as well as noise,and the segmentation is not accurate when dealing with the gray uneven image.Compared with this method,convolutional neural network can segment more complex medical images,however,it needs a lot of artificial marks.The feature extraction mechanism of the network leads to its inaccurate decisionmaking of atypical boundary features.In this paper,the fusion method of convolutional neural network and active contour model is studied for medical images with small training set.The main research work is as follows:(1)The basic methods of active contour model and convolutional neural network are summarized.Some classical models and network structures are introduced.In addition,the advantages and disadvantages of each classical segmentation method are briefly summarized.At the same time,the advantages and problems of the active contour model and convolutional neural network segmentation method are analyzed.(2)In view of the problems of uneven gray level,unclear edge and initial sensitivity of active contour model,a new energy function is proposed by combining convolutional neural network method,which is composed of fitting energy term,fidelity term and regular term.First,the prior information is obtained by unet 3+ network.The prior information is used as the fitting energy term to construct a new energy function,which minimizes the energy function and constrains the curve evolution.The new energy function can effectively improve the robustness of the model to noise and initial contour.Meanwhile,edge stop function is added to the energy function as the weight of fitting energy term and fidelity term,and the edge information of image is introduced to supplement the edge features weakened by neural network in the training process,as well as,the segmentation results are optimized.(3)In view of the problem that the method of active contour model needs to be adjusted manually,a fully automatic segmentation model based on convolution neural network and active contour model is proposed.This model is modified on the basis of Deep Active Lesion Segmentation(DALS),in which the information of each layer in the model's down-sampling stage is transferred to each layer in the up-sampling stage,so that the model can make full use of the characteristic information and the full-scale information of each stage.In the meantime,the dense block and multi-scale dilated module are used to narrow the network,reduce the number of feature graphs,reduce the amount of computation that increases as connections are added,and improve the calculation efficiency.The method first obtains the estimated probability map by network model,then calculates the initial distance function and two parameter variables of the active contour model according to the probability graph.Finally obtains the segmentation result by minimizing the energy function.In conclusion,this paper systematically introduces the methods of convolution neural network and active contour model,meanwhile,studies the fusion of the two methods.Two image segmentation methods with combining convolutional neural network and active contour model are proposed.The experimental comparison between the two methods is carried out on skin lesions and chest X-ray images with the existing segmentation methods,which verifies the effectiveness of the proposed model.