Multi-label Bone Segmentation Method Based On Multi-model Fusion For Chest And Abdomen CT Images

Chest and abdomen are the high incidence of human diseases.CT images of chest and abdomen can be used to diagnose bone and blood vessels.It is important to design an automatic segmentation method to segment a variety of bones in chest and abdomen CT images.The segmented bones can be directly used in CT bone analysis projects,which can serve for 3D scan diagnosis of bones and orthopedic surgery planning.It can also be used for the bone removal operation in CT vascular analysis project,which can meet the requirements of vascular display and location.At present,the accuracy of traditional image segmentation methods for thoracic and abdominal CT image bone segmentation is not high,and it is difficult to subdivide the bone categories.The deep learning method can automatically extract image features and achieve high segmentation accuracy.Therefore,according to the characteristics of chest and abdomen CT images,a bone segmentation method based on multi-model fusion is proposed.This method can divide the bones in chest and abdomen CT images into 13 categories.Firstly,the data is preprocessed to reduce the interference of irrelevant factors.After that,the data will be sent to three models for training.The three models are two improved 2D models based on 2D-Unet and one improved 3D model based on 3D-Unet.Two optimization strategies,residual module and attention mechanism,are added to the three models.Two2 D models were trained and predicted with axial and coronal data respectively.When the data is predicted,three prediction probability maps are obtained by using three models.The three prediction probability maps are fused into one probability map as the final prediction result.Then the corresponding label data is generated according to the fusion result.Finally,the label data is processed to get the final segmentation result.The experimental results were evaluated by dice similarity coefficient and sensitivity.The experimental results show that the 2D-Unet and 3D-Unet with residual module and attention mechanism have better segmentation effect than the original network model and the model with single optimization strategy.After the fusion of the three improved models,higher segmentation accuracy can be obtained.