Research On Liver Image Segmentation And Classification Technologies Based On Convolutional Neural Network

With the maturity of imaging technology,medical imaging plays an especially important role in the auxiliary medical diagnosis,especially for the diagnosis of abdominal diseases.It can present the internal condition of abdominal cavity to help doctors make more effective diagnosis.However,it is still a great challenge for accurately diagnosis because that there is little density differences between different organs and tissues in those images.Therefore,it is of great and far-reaching significance to do research on computer image processing technologies for auxiliary image diagnosis.This paper carries out research of medical image processing through machine learning,deep learning and other related technologies which have made remarkable achievements in natural image processing based on abdominal CT.1.To solve that it is difficult to get accurate segmentation for small areas in liver and liver lesion segmentation,this paper proposes a model named IsU-net to achieve segmentation of livers and meanwhile,using transfer learning method to achieve the segmentation of liver lesions.In proposed IsU-net,firstly,in down-sampling processing,different convolution kernels are used in different layers to extract more multi-scale features.Secondly,in up-sampling,a selecting gate is designed to screen more favorable features from corresponding down-sampling as complementary features to reduce the loss of information.Finally,during the training process,losses of all stages in up-sampling are comprehensively considered to optimize the network and promote the performance.In addition,due to data scarce when training liver lesion segmentation network,parameters of the liver segmentation model are transferred to reduce the consumption in training.Experiments show that the proposed IsU-net can effectively alleviate the problem of small area segmentation.2.To solve that it is difficult to diagnose some liver diseases because that there is no obvious density differences between normal liver tissues and pathological liver tissues,and among different types of pathological liver tissues in plain CT.Meanwhile,there is less effective training data.This paper proposes a dual flow network model combined with multiple transfer learning to achieve classification of four kinds of pathological liver tissues.In this network,firstly,to realize feature supplement and enhancement,a double flow model is designed to do feature extraction.Secondly,an ensembled classifier is designed to obtain more accurate and reliable classification results,and a dual constraint is proposed to optimize the network.In addition,on the one hand,due to lack of data,parameter-based transfer learning is used to assist the target network initialization;on the other hand,in order to alleviate the over fitting on testing data,domain adaptation is adopted during training process to improve the model performance.Experimental results show that the proposed model has better classification effect and stronger fitting ability.