Pancreas Segmentation Algorithm Of CT Image Based On Enhanced Semantic Information

With the rapid development of artificial intelligence technology,Computed Tomography(CT)technology can image the location,shape and organization of pancreas organs,which is an important premise for accurate localization of lesions,quantitative analysis for doctors,and research on anatomical structures.Through the high-efficiency image processing capability of the computer,the accurate and efficient segmentation of pancreas organs from CT images can not only assist clinical disease diagnosis but also play a very important role in the formulation of tumor ablation surgery.By analyzing the characteristics of clinicians’ cognition of pancreas organs in CT images.This thesis studies two focuses,the 2D segmentation algorithm and 3D segmentation algorithm,and proposed a CT image pancreas segmentation algorithm based on enhanced semantic information.The main contribution contents includes:(1)To enhance the ability of the 2D segmentation algorithm to extract pancreas semantic information from CT images and improve the accuracy of pancreas segmentation,this thesis proposes a pancreas segmentation algorithm based on variational autoencoders to enhance semantic information.First,the variational autoencoder is introduced to enable the network to reconstruct deep semantic features,and the reconstruction of the input images can effectively improve the problem of information loss caused by downsampling;second,the skip connection is improved through the attentional feature fusion module,so that the network pays more attention to the weak and small pancreas organ area.The experimental results show that the algorithm achieved the highest DSC of 90.15%,and it is verified by ablation experiments that adding the variational autoencoder module and the attention feature fusion module at the same time achieved the best segmentation results.(2)The 2D pancreas segmentation algorithm can obtain more accurate and rapid slice-byslice segmentation results,so it has always attracted much attention.Although the average accuracy of 2D segmentation results is high,there is a lack of contextual information between slices,and the segmentation accuracy of each slice is different,resulting in poor segmentation continuity between slices.This leads to decrease performance of surface smoothness and continuity when obtaining 3D pancreas organs.Therefore,this thesis proposes a pancreas segmentation algorithm based on improved V-Net,which fuses the information within and between slices,effectively improves the surface continuity and smoothness of the 3D model,and improves the authenticity of the prediction model.The experimental results show that the algorithm achieved the highest DSC of 85.52%,and the effectiveness of the attention module is verified by ablation experiments,and the use of the DSC loss function can make the network model achieve the highest DSC.(3)To overcome the problem of under-segmentation or over-segmentation between the 3D pancreas organ segmentation results and the original label,and further improve the accuracy of the segmentation of the 3D pancreas model,shape constraints are added on the basis of the algorithm in Chapter 3,and the skeleton prediction branch of the pancreas is added to achieve the goal of multi-task learning,together with location information,optimizes the training of network models.The experimental results show that the highest DSC of the algorithm achieved 86.21%,and effectiveness of adding shape constraints is verified by ablation experiments,and the phenomenon of under-segmentation or over-segmentation can be effectively suppressed.Although in the existing 3D segmentation algorithm,the average network performance of the algorithm is lower than that of the 2D segmentation algorithm,the application of the 3D segmentation algorithm can effectively improve the continuity and smoothness between slices,which can provide a more real and reliable basis for doctors to diagnose diseases.