Low-dose CT Based Whole Body Adipose Tissue Quantification Analysis

The content of adipose tissue in the human body is closely related to clinical indicators such as the incidence of many diseases.It plays a decisive role in the design of diagnosis and treatment plans,prediction of life cycle,and estimation of rehabilitation conditions.Therefore,accurately quantify body adipose tissue is pathologically significant.Because of its non-invasive scanning method,low-dose CT has the characteristics of less radiation and easy access in daily diagnosis,and is increasingly used in the study of quantitative analysis of human body fat content.Due to the different roles played by different adipose tissues and their different influences on clinical indicators,it is necessary to quantify them separately.However,there is currently no unified definition of different adipose tissues under CT images.In addition,the distribution of whole body adipose tissue is very complex and the distribution range covers the entire human body.All these present obstacles to accurate quantification of adipose tissue.In order to overcome the above difficulties,this paper firstly uses three customized contours: the skin area(Skn),the outer aspect of skeletal musculature(OAM)and the inner aspect of skeletal musculature(IAM),which help to clearly divide the adipose tissue in the human body into subcutaneous adipose tissue(SAT),intermuscular adipose tissue(IMAT)and visceral adipose tissue(VAT).This paper established a whole body adipose tissue definition system and a body adipose tissue data including 50 low-dose CT scans under the guidance of the system set.In addition,this paper proposes a contour-based dual-decoder attention segmentation network,which fully extracts the prior knowledge of the anatomical structure contained in the three contour regions,and embeds the segmentation of the contour region into the network structure as an auxiliary branch to promote the accurate segmentation of adipose tissue in low-dose CT images.At the same time,a loss function that can pay more attention to the manual correction of labels in the semi-automatic labeling process is designed,so that the segmentation network pays more attention to complex regions that are prone to mis-segmentation,so as to improve fat Segmentation accuracy of the tissue.In this paper,the proposed algorithm is verified on the established data set.The experiments show that the Dice coefficients of the proposed model on three adipose tissues of SAT,VAT and IMAT are 94.64%,92.08% and 81.72%,respectively.Compared to the classic medical image segmentation model UNet,the Dice coefficient is improved by 1.76%and the proposed algorithm has better stability.In addition,this paper also compares the difference of labels between different annotators,which proves to a certain extent that the upper limit of the Dice coefficient that can be achieved by automatic segmentation of SAT should be about 97.89%,indicating that the accuracy of the proposed model is only about3% away from the upper limit.The gap further verifies the effectiveness of the algorithm in this paper.