Multi Label Segmentation Of Bone Tissue In Chest And Abdomen CT Images Based On U-shaped Convolutional Network

At present,in clinical diagnosis,doctors often use computed tomography(CT)to diagnose the patients’ chest and abdomen bone diseases.Doctors need to automatically identify different types of bone tissue to improve the efficiency of diagnosis.A multi label segmentation method of bone tissue in chest and abdomen CT images based on U-shaped convolutional network is proposed.The proposed method of chest and abdomen bone segmentation is implemented in nn UNet.According to the characteristics of dataset used in the experiment,the medical image segmentation scheme provided in nn U-Net is selected.Firstly,the training data is preprocessed by cropping,resampling and data standardization.Then,the dataset is augmentation by many schemes,but the mirror augmentation of nn U-Net will make the symmetrical bones lose the left and right position information,and cause the confusion between the left and the right,so the mirror augmentation is removed in the experiment.In the network training stage,based on 2D U-Net and 3D U-Net provided by nn U-Net,3D Attention U-Net with attention mechanism and 3D Res U-Net with residual connection are added,and then trained the four network models.These network models are evaluated by a variety of evaluation methods.3D Res U-Net can be designed deeper and learn more image features because of the residual connection added to the network,and it is the best performance among the four network models.In the prediction stage,the network model is used to predict the test data.The results show that there are holes in the left and right scapula of some data.In the post-processing,the holes are filled by morphological closed operation.The experimental data comes from the hospital,and the segmentation targets of these data are manually labeled to obtain the experimental dataset.29 sets of data were used to verify the proposed multi label segmentation method of bone tissue.The results show that the method can segment 11 kinds of bones accurately,and the average dice coefficient of the test data can reach 0.96.