Research On CT Image Liver And Tumor Segmentation Algorithm Based On Deep Learning

Liver cancer is now one of the most common cancers worldwide,with high incidence and mortality rates,which seriously threatens people’s life and health.Currently,computed tomography(CT)has become the most commonly used medical imaging method in the diagnosis and treatment of liver cancer,and CT-based computer-aided liver and tumor segmentation is a very crucial guide for preoperative planning,intraoperative reference,and postoperative evaluation.However,characteristics of liver tumors such as location,number,size,and shape vary from person to person and are highly heterogeneous.Segmentation algorithms based on image processing or traditional machine learning are usually difficult to effectively extract and utilize the high-level semantic information in CT images,and the accuracy of segmentation results of diseased liver and tu-mors is very limited.Compared with the former algorithms,the popular deep learning algorithms can bring significant improvement in the accuracy of liver and tumor segmentation,but the segmentation results of the complex details in CT images are still not satisfactory,and the disadvantages such as large storage space occupation and high computational complexity also limit the popular deep learning methods to large-scale clinical applications.To address the problems above,two lightweight and high-performance networks for CT image liver and tumor segmentation are proposed in this thesis,and the main work of this thesis is as follows:(1)There are a lot of parameters and computation in popular 3D fully con-volutional networks including V-Net,which make them difficult to convergence.To address this issue,we proposed a lightweight V-Net with decoder deep super-vision mechanism.The main architecture of encoder and decoder is constituted by inverted residual bottleneck blocks(IRBB),and adjacent encoder or decoder modules are connected by lightweight downsampling or upsampling module.These designs heavily reduced the parameters and computation of the proposed network,meanwhile,the expansion path and constrained path of IRBB enhances the network’s feature extraction and selection capacities.The deep supervision module following decoder module can inject gradient value into the internal lay-ers and improve the quality of feature maps in the backpropagation phase,which can increase the segmentation performance of the proposed network as well.Experiments show that the proposed network can attain a better liver and tumor segmentation accuracy with lower storage and computational cost than other popular 2D or 3D networks.(2)The popular CT image liver and tumor segmentation networks are not able to effectively utilize global context information,which leads to bad seg-mentation results of complex boundaries or internal regions of liver or tumor.To address this issue,we proposed a U-Net with dual attention module and du-al-path decoder.We use MobileNetv2 to build the encoder of the proposed net-work and use IRBB to build the main architecture of the decoder,which makes this network highly lightweight.The dual attention module can help the network to capture channel-dimensional and spatial-dimensional contextual information globally and enables the network to model the interdependencies between fea-tures regardless of their distance,enhance the response of regions of interest in the feature map and suppress the noise appearing in irrelevant regions by means of weighted sums,which can effectively improve the continuity and consistency of the network’s understanding of features in the internal regions of the liver or tumor.The dual-path decoder provides the network with much more differential information,which helps to improve the sensitivity and accuracy of liver or tu-mor boundary discrimination.Experiments show that compared with popular networks,the proposed network can attain a much better liver and tumor seg-mentation accuracy,meanwhile the storage and computational cost are minima.