Research And Application Of Abdominal CT Image Segmentation Based On Dense Connection And Multi Branch Structure

Medical image segmentation plays an important role in computer-aided medical diagnosis,organ location,3D reconstruction and operation planning.Liver is an irreplaceable important organ in human body,which maintains the life activity and metabolism of human body,and liver tumor is a kind of malignant tumor with high mortality.Therefore,it is of far-reaching significance to study an accurate automatic segmentation algorithm of liver and liver tumor for clinical diagnosis and surgical decision-making.Abdominal multi organ segmentation can not only improve the efficiency of segmentation algorithm,but also improve the robustness of segmentation results by using the relative stability of spatial position relationship between abdominal organs.Therefore,accurate abdominal multi organ automatic segmentation can help doctors diagnosis more efficiently,which has high research value.CT image is widely used because of its high efficiency,high cost performance and high resolution.In this paper,the abdominal CT image is taken as the research object,and the algorithms of liver and liver tumor segmentation and abdominal multi organ segmentation are studied.The main research contents are as follows:(1)In order to solve the problem of unbalanced samples and under segmentation of unconnected regions caused by the low proportion of liver cross-section in axial images,a method of liver and liver tumor segmentation based on dense connected U-Net model is proposed.This algorithm introduces a dense connection structure in the U-Net model,combines the low-level edge features with the high-level semantic features,alleviates the information loss in the process of feature transfer,and can better extract the semantic features of the target organ.In the upper sampling part,the threshold selection mechanism is used to make the target organ boundary smoother after resolution reduction.A new loss function is designed to solve the problem of sample imbalance.Experiments show that the segmentation results of DCU-Net model have higher accuracy and robustness.(2)In order to solve the problems of large spatial differences between different organs in multi organ segmentation and competition in the training process,a multi organ segmentation network model(MFRNet)based on multi branch and feature recalibration structure is proposed.The convolution kernel of multiple channels in the multi branch structure can extract all the features of multiple organs from multiple angles and improve the feature diversity of the network.Through the feature recalibration structure and attention mechanism,the feature map is screened from both the space and the channel to obtain the feature map which is helpful for segmenting the target.The experimental results show that MFRNet can improve the accuracy of multi organ segmentation.(3)The abdominal CT image segmentation system is designed and developed,which realizes the functions of abdominal CT image import,preprocessing,segmentation and optimization.The combination of liver and liver tumor segmentation model and multi organ segmentation model with the system can alleviate the work intensity of doctors’ manual segmentation,which is of great significance to the development of intelligent medical treatment.