Research On Liver Tumor Segmentation For Enhanced CT Based On 3DA-U-Net Framework

Liver cancer is a disease with high morbidity and mortality worldwide.The use of image analysis technology to detect lesions in the liver as soon as possible is very important for the diagnosis and treatment of liver cancer.However,due to the complex connection between the liver and other organs,the shape,location,scale and texture of liver tumors vary from person to person.The segmentation of liver tumors has always been a difficult problem in medical image segmentation.With the development of artificial intelligence technology,the use of deep learning to achieve effective liver tumor segmentation has become a focus of attention.Based on the previous research foundation of the research group,this thesis establishes the 3DA-U-Nets framework,and proposes a 3DA-Dense U-Nets-LSTM enhanced CT liver tumor segmentation method based on ensemble learning.This method increases the attention mechanism of the segmentation network for key channels and key spaces,and uses dense connection modules and recurrent neural networks to optimize the convolutional network structure to enhance the extraction ability of the network.Finally,the ensemble learning is used to fuse the enhanced CT three-phase image features.The main work and results of this thesis are as follows:(1)A novel unified framework for 3DA-U-Nets-enhanced CT liver tumor segmentation is proposed.A new three-dimensional attention module is established on the basis of the traditional attention mechanism.By scanning the global image,this module focuses on the region and characteristics of the target,reduces the attention of information irrelevant to the tumor,and improves the efficiency and accuracy of the liver tumor segmentation network.The module is fused with the U-Net network to obtain the 3DA-U-Nets framework,and then the network cascade method is introduced.The first-stage network generates the liver mask,and the second-stage network performs the segmentation of liver tumors.By increasing the target attention,the goal of improving tumor segmentation accuracy is achieved.(2)A 3DA-Dense U-Nets-LSTM liver tumor segmentation method is proposed.In order to make better use of the sequence of CT images and improve the problem that the U-Net network is difficult to obtain higher segmentation accuracy due to insufficient depth,the convolutional layer is replaced with dense connection blocks on the basis of 3DA-U-Nets,and then the bidirectional C-LSTM module is added to effectively extract the information between adjacent slices,so as to deepen the network depth,extract data sequence,improve network segmentation accuracy,and speed up network convergence.(3)A Stacking-3DA-Dense U-Nets-LSTM liver tumor segmentation method is proposed.Considering the imaging characteristics of enhanced CT,the three-phase images of different tumors contain different tumor information and have different effects on the results.The ensemble learning method is used to fuse the feature information of three-phase enhanced CT images,and the stacking combination strategy is used to effectively utilize multimodal medical images and achieve the purpose of further accurate and efficient segmentation of liver tumors.In this thesis,the performance of the proposed algorithm is deeply analyzed through experiments.Due to the lack of effective enhanced CT datasets in commonly used public data,we cooperated with Jiangsu People’s Hospital and obtained enhanced CT data of 40 patients as experimental datasets,each of which was performed by professional physicians.The comparison networks used in the experiments are FCNs network,U-Net network,SE-U-Net network and H-Dense UNet segmentation network,etc.The experimental results verify the superiority of the performance of the proposed algorithm.