Body Composition Tissue Segmentation On CT Image

Body composition analysis(BCA)plays a crucial role in the diagnosis and treatment of many diseases,like obesity and type 2 diabetes mellitus,and health management.Computed Tomography(CT)imaging technique makes it possible to visualize the human anatomy and to quantify body composition tissues for individuals precisely.But this method requires annotations of related body composition tissues that are annotated by professional experts,which becomes an important limitation for the application of such methods.This reaearch aims to automatically segment and quantify body composition tissues on CT images using deep learning technique.Different body composition tissues are located in different body regions with ambiguous boundaries and are hard to distinguish.To improve the automatic segmentation performance,firstly,this paper defines body regions,where each key body composition tissue is contained in one specific minimal bounding body region.Secondly,this paper devises Geographic Attention Network(GA-Net)based on the body region and body tissue models,which provides rich spatial prior information to the network.Because body regions contain rich semantic and spatial information that is highly related to the target body composition tissues,GA-Net can significantly improve the accuracy of identifying these body composition tissues by leveraging the information.Particularly,GA-Net is based on a novel dual-decoder schema that is composed of the body tissue decoder and the body region decoder.The body tissue decoder segments the target body composition tissues and the body region decoder predicts the related body regions,as an auxiliary task.The feature maps of body regions and body tissues are fused through the soft attention mechanism to obtain the region attention information related to the body tissues.The experiments are conducted on a private dataset that includes 50 low-dose unenhanced PET/CT images for segmenting four essential body composition tissues,including subcutaneous adipose tissue,visceral adipose tissue,skeletal muscle tissue,and skeleton,automatically.The Dice coefficient,Hausdorff Distance(HD),and 95% HD(HD95)are used for evaluation.The results show that GA-Net achieves superior Dice coefficient results,and improves HD and HD95 metrics in most of the comparison experimental groups.