Research On Medical Image Segmentation Method Based On Improved Convolutional Neural Network

At present,we are in the era of big data,all kinds of data are increasing,and data processing has become a driving force of social development.And at the same time,with the advance of science and technology,artificial intelligence can be seen everywhere in life.For example in the great medical treatment,the automation of image segmentation is a significant divide of the development of intelligent medical system.Doctors treat patients through automated medical image segmentation to improve medical level and efficiency.This way is bound to become a development trend in the near future.However,compared with other images,medical images have some particularities,so the task of medical image segmentation has certain difficulties and challenges.Therefore,the research on medical image segmentation methods is very important.This article mainly studies on the task of medical image segmentation.In order to combine medical image segmentation with artificial intelligence,we study many of CNN's structures,understand the fundamental knowledge of them,and know the development and applications of convolution networks.Then we compare and analyze the structures and the key points of different algorithm models,at the same time,some traditional image segmentation methods are also compared with convolution neural network.Then,for the problems existing in some medical image segmentation tasks,based on the U-net network,some schemes are designed to improve the network model in order to improve the performance of medical image segmentation.The main contents of the proposal are as follows:(1)According to channel attention and U-net,we design a medical image segmentation model.So as to deal with the question that it is too hard to focus on the useful and significant information features in image segmentation methods,inspired by the attention mechanism,a channel attention mechanism based on U-net network is proposed.The channel attention mechanism module is added between convolution and pooling,and the feature map obtained after the image passes through the convolution layer is processed by the module,and each feature channel is given a corresponding weight.This weight represents the degree of correlation between this channel and important features,the smaller the weight is,the more uncorrelated the channel is,and the larger the weight is,the greater the correlation is.Then the weight is multiplied with the input feature channel by channel to get the final weighted feature.By using this channel attention mechanism,the useful information is weighted reasonably.It can selectively emphasize information and suppress irrelevant redundant information,so as to improve the segmentation accuracy of the network model.Then we design experiments to verify it,through the experimental verification in the data set of liver and lung,the Dice similarity coefficient of the proposed method is 2.7% and 1.8% higher than that of the original U-net model,respectively.So it proves that the proposed method can segment medical images better.(2)This paper also proposes a medical image segmentation method based on self-attention mechanism and residual error.Because medical images often have some irregular areas,and the characteristics of these areas will depend on the information corresponding to certain positions in the global image.Therefore,it is necessary to increase the perceptual field of the feature map to obtain more global context information.To solve this problem,a self-attention mechanism for medical image segmentation is proposed,which calculates the association between each element and other elements,and then uses the calculation to obtain a weighted representation.At the same time,the residual structure is used to combine the features of the input image with the features of the self-attention output to obtain the enhanced features,so as to retain the spatial information of the features.Similarly,we designed the verification experiments on two datasets,and it is worth mentioning that in training we used the cross-entropy loss function to optimize the network.The experimental results show that in the data set of two different organs,our method is superior to the original algorithm,so the method is proved to be successful and effective.(3)Next,this article proposes another method,that is a medical image segmentation technology based on multi-scale feature extraction.Medical images sometimes have different texture shapes at different scales,and the original model often uses only one receptive field in the feature extraction process,which makes the extracted features relatively single,which will also have a certain impact on the segmentation accuracy.Therefore,this paper proposes to add a multi-scale feature extraction module on the basis of the U-net network model.The receptive field is added to the feature extraction of each layer of the network,and the hierarchical residual connection constructed in a single residual block,To capture multi-scale features at a finer-grained level,and then send a thicker feature map that incorporates multiple scale information to the next layer of the network to optimize the network feature extraction process,so as to achieve the purpose of improving segmentation accuracy.The feasibility and effectiveness of this method are proved through experiments set in two organ data sets.