Research On Liver And Liver Tumor Automatic Segmentation From CT Images Based On Deeply Supervised Network

Liver tumor is the sixth most common cancer and the third leading cause of cancer death.Early diagnosis of liver tumor can assist radiologists in making appropriate treatment plans.In clinical practice,the contour of liver and liver tumor is delineated by radiologists,which is time-consuming and subjective.Liver and its surrounding organs have similar gray levels,and liver tumors have a wide variety of locations,volumes and shapes.Based on the above problems,the Hierarchical Contextual Cascaded Fully Convolutional Network(HC-CFCN)model is proposed to segment livers in CT images and the Spatial Feature Fusion Network(SFF-Net)model is proposed to segment liver tumors in CT images.(1)The hierarchical contextual information can improve the precision of the automatic segmentation method in medical image segmentation tasks.However,the existing convolutional neural network ignores the importance of hierarchical contextual information.To fully extract more hierarchical contextual information in CT images,we use the HC-CFCN model to segment livers in CT images.The model utilizes a dual network structure and a multi-channel strategy to explore inter-layer context information.Besides,a method named energy map is applied to our model to segment the liver efficiently.Experimental results on the MICCAI 2017 Li TS dataset show that the segmentation accuracy of theHC-CFCN model is higher than that of the U-Net,FCN+3DCRF,and V-Net models.(2)Due to the lack of position information and shape information,accurate segmentation of liver tumors depends on the extraction of spatial information.The existing convolutional neural network loses a lot of spatial information in the down-sampling phase.Besides,the feature fusion is difficult.To extract more spatial information,we present the SFF-Net model segments liver and liver tumors sequentially.We add skip-connections to deeply supervised network between side-output layers and up-sampling layers.Besides that,we utilize a feature fusion block to merge different feature maps.On MICCAI 2017 Liver Tumor segmentation challenge test data,our model achieved a Dice Global score of 0.592,Dice per case score of 0.746 and tumor detected precision of0.369.Ablation experimental results on the MICCAI 2017 Li TS dataset show that the accuracy of segmentation increased from 0.090 to 0.369.