3D Body Composition Segmentation in Abdomen and Pelvis CT using Subdivided Labels and Random Patch

Minyoung Kim, Ji-Won Kwon, Kwang Suk Lee, Taehoon Shin

Keywords: random patch, 3D, body composition, segmentation, APCT, Swin UNETR

Abstract: The distribution and volume of fat and muscle in APCT play an important role as a biomarker. In this study, APCT data from 200 individuals who underwent health screening was labeled into three classes of fat and four classes of muscle. Based on this labeling, 3D patch-wise segmentation was performed by Swin UNETR on the whole abdomen and pelvic scan images. The test results showed an overall class average of 0.9227 DSC. This study conducted 3D whole-abdomen body composition segmentation using a total of eight segmented body composition labels including the background and verified its feasibility using random patches effective for the data and task.