Research On Liver CT Image Segmentation Method Based On Deep Learning

Liver cancer is currently one of the most common cancers in the world,posing a serious threat to human life safety.In the process of diagnosis and treatment of liver cancer,it is very important to accurately and quickly segment the liver and liver tumor area.Finally,accurate positioning and identification of the liver tumor contour is of great significance for doctors.Computed Tomography(CT)is one of the most important methods for the auxiliary diagnosis of liver tumors,which can effectively improve the treatment rate of liver tumors.Liver tumor segmentation by traditional image-based and machine learning-based methods requires high professional level and clinical experience of doctors,which is inefficient.In addition,traditional segmentation methods are easily affected by the subjective judgment of doctors during manual intervention.The automatic liver CT image segmentation method based on deep learning shows better performance and higher efficiency than traditional image-based and machine learning methods.It can integrate the two stages of feature extraction and tumor segmentation into one to improve performance and reduce redundant steps.The current deep learning segmentation methods have some problems in the task of liver tumor segmentation,such as insufficient accuracy,fuzzy boundary regions,and undersegmentation of small target regions.In this regard,this paper carries out research from the following aspects:(1)In this paper,a liver CT tumor segmentation method based on Collaborative Attention UNet(CA-Unet)was proposed.The CA-Unet has the basic structure of encoder and decoder.By adding collaborative attention mechanism blocks on the residual paths of every two convolutional layers of the encoder,it can better extract the local feature information of the image,and can pay attention to the spatial position information of the region of interest of different levels of feature images.At the same time,Dy Relu activation function is introduced to enhance the representation ability of the model,which is significantly helpful for the accurate segmentation of liver and liver tumors.In addition,the Focal Tversky Loss function was introduced into the network training process to greatly reduce the influence of false positive and false negative samples on the training.A large number of experimental results on public datasets and hospital-made datasets show that CA-Unet can segment liver and liver tumor regions more accurately than a variety of advanced segmentation models.(2)A liver CT image segmentation method based on Dual-path Self-attention Multi-scale Feature Fusion Network(DS-MSFF-Net)was proposed.Specifically,DS-MSFF-Net uses the first branch path to extract detailed semantic feature information and the second branch path to extract global semantic feature information.The feature information obtained through these two paths are mutually compensated and fused,and the high-precision segmentation results are finally obtained.This architecture solves the problem of incomplete extraction of semantic feature information from large resolution CT images.Extensive experiments on Li TS2017 public dataset and hospital-made dataset have shown that DS-MSFF-Net performs well in liver and liver tumor segmentation tasks.(3)Design of liver CT auxiliary diagnostic software.The software has several main functions,including liver CT image selection,image segmentation,calculation of lesion area,calculation of segmentation accuracy,and segmentation time calculation.The graphical user interface of the software was designed using Pyqt5 and Qt Designer.By encapsulating the DS-MSFF-Net algorithm proposed in Chapter four of this paper,the software can perform high-precision segmentation of liver and liver tumors,and then realize the function of auxiliary diagnosis.