Research On Liver Tumor Image Segmen-tation Based On Convolutional Neural Network

Liver tumor is a disease that seriously endangers human life and individual health.Liver malignant tumor is one of the causes of cancer death.At present,one of the main treatments for some types of liver cancer is the removal of liver tumors.According to CT imaging(CT),magnetic resonance imaging(MRI),and other medical imaging technology can achieve high resolution imaging of tumor pathological changes,provideing crucial information for treatment planning,however,based on the image of the tumor diagnosis and treatment are dependent on the doctor’s subjective judgment and operation.With the increasing amount of data and information,the doctor manual analysis of tumor outline time-consuming and laborious.Therefore,how to realize automation of CT film accurate segmentation has high clinical value.In this paper,the deep learning method was used to study the problems in CT liver tumor images,such as small tumor proportion,difficult segmentation,multiple small tumors can not be effectively segmented,and the boundary of segmentation is fuzzy.Due to the small proportion of liver tumors in abdominal CT images,the two-stage segmentation method was adopted in this paper to limit the results of tumor segmentation to the liver region.First,the liver region was segmenting,and the obtained liver was used as the region of interest,excluding the influence of other abdominal tissues on tumor region segmentation and reducing false positives.When the liver image segmentation,this paper will be the most popular in the field of medical image segmentation U-.net as a benchmark network,add dual attention mechanism in network module,the module is made up of channel attention yourself module and location module in parallel,through build rich context dependencies on local characteristics,to enhance useful features,so as to get more accurate segmentation results.Liver segmentation results obtained by the second step of cutting processing,improve the efficiency of segmentation,image for small tumors,we in the liver segmentation model on the basis of the fusion feature pyramids multi-scale prediction,the characteristics of the model in the decoder path characteristics of each layer figure connected,make its depth is equal,and sample up to a fixed dimension,the characteristics of this set of figure as semantic segmentation layer input,strengthen the semantic segmentation advantage;Then the residual module and batch group normalization were used to optimize the improved network structure.Then,Focal Loss and Tversky Loss mixed loss functions were used to solve the problem of unbalanced data categories and the poor ability of network to identify difficult tumor samples.The experimental results show that the network model proposed in this paper can better segment the liver and liver tumor region,which not only has a great improvement in accuracy,but also has a good effect for the detection of multiple small tumors and the detection of tumor edge.The performance of the method proposed in this paper in the DICE coefficient evaluation index is comparable to that of the excellent algorithms published in the current conference,and it can provide certain clinical reference value.Finally,this paper uses QT5 to design an image segmentation software for liver tumors,which can read NIFTI(Neuro Imaging Informatics Technology Initiative)medical CT images and display the segmented tumor regions,so as to provide auxiliary reference for doctors in radiotherapy of liver cancer.