The 2D-3D hybrid mean-teacher network achieved high mean Dice similarity coefficients for subcutaneous adipose tissue (0.989), visceral adipose tissue (0.979), and skeletal muscle (0.973).
A novel 2D-3D hybrid mean-teacher network can accurately segment abdominal body composition from CT scans using only 10% labeled training data, achieving comparable performance to manual segmentation.
Abstract Background Human body composition is a key indicator of metabolic health, cardiovascular risk, and other health conditions. Segmentation of different adipose tissues and skeletal muscle based on CT images is a commonly used way to assess body composition accurately. Purpose Most current body composition analysis methods focus on single cross‐section CT images, which may have a bias of selecting single‐slice images to represent whole body composition. Existing 3D deep‐learning‐based segmentation methods usually require densely annotated training volumetric data and face challenges of intra‐ and inter‐variability among subjects. This article aims to realize accurate volumetric abdominal wall segmentation and body tissue extraction from CT scans with sparse‐labeled training data. Methods This study collected 70 CT scans from one public dataset and one private dataset, which were divided into the training (public, ), internal test (public, ) and external test (private, ). About 10% of the slices in training scans are labeled to develop the model and all test scans are delineated slice‐by‐slice by radiologists. A novel semi‐supervised segmentation method based on a 2D–3D hybrid mean‐teacher network (2D‐3D HMT) and an automated body composition assessment pipeline is proposed. Specifically, a 2D teacher model trained with fully‐labeled slices guides the fine‐tuning of two 2D and 3D student models, taking advantage of the volumetric model while avoiding intensive slice‐by‐slice labeling work. In addition, an active contour regularization is incorporated into 3D model to improve boundary smoothness. Then, different adipose and muscle tissues are extracted based on the predicted abdominal wall and other ROIs using specific tissue HU ranges. Results Experiments were conducted on the internal and external datasets to evaluate the semi‐supervised segmentation model and body tissue assessment pipeline. Additionally, the public large‐scale AATCT‐IDS dataset with adipose tissue annotation was also used to evaluate the proposed method. The proposed method achieved high mean Dice similarity coefficients (DSC) on the internal dataset (0.989, 0.979 and 0.973) and external dataset (0.986, 0.982 and 0.973) for subcutaneous adipose tissue (SAT), visceral adipose tissue (VAT) and skeletal muscle (SM), respectively. The Pearson's correlation analysis was conducted and showed the volumes of SAT, VAT and SM by automated segmentation are strongly correlated with those of manual segmentation (, , , , , , respectively). Compared to related body component assessment methods, the proposed method reached comparable performance while only using 10% labeled data. Conclusions The proposed semi‐supervised segmentation method and pipeline can effectively perform body composition assessment with very sparse‐labeled CT images.
Hu et al. (Fri,) conducted a other in Body composition assessment (n=70). 2D-3D hybrid mean-teacher network (2D-3D HMT) vs. Manual segmentation was evaluated on Mean Dice similarity coefficients (DSC) for subcutaneous adipose tissue, visceral adipose tissue, and skeletal muscle. The 2D-3D hybrid mean-teacher network achieved high mean Dice similarity coefficients for subcutaneous adipose tissue (0.989), visceral adipose tissue (0.979), and skeletal muscle (0.973).