Annotation-Efficient Strategy for Segmentation of 3D Body Composition

Body composition as a diagnostic and prognostic biomarker is gaining importance in various medical fields such as oncology. Therefore, accurate quantification methods are necessary, like analyzing CT images. While several studies introduced deep learning approaches to automatically segment a single slice, quantifying body composition in 3D remains understudied due to the high required annotation effort. This study proposes an annotation-efficient strategy using an iterative self-learning approach with sparse annotations to develop a segmentation model for the abdomen and pelvis, significantly reducing manual annotation needs. The developed model demonstrates outstanding performance with Dice scores for skeletal muscle (SM): 0.97+/-0.01, inter-/intra-muscular adipose tissue (IMAT): 0.83 +/- 0.07, visceral adipose tissue (VAT): 0.94 +/-0.04, and subcutaneous adipose tissue (SAT): 0.98 +/-0.02. A reader study supported these findings, indicating that most cases required negligible to no correction for accurate segmentation for SM, VAT and SAT. The variability in reader evaluations for IMAT underscores the challenge of achieving consensus on its quantification and signals a gap in our understanding of the precision required for accurately assessing this tissue through CT imaging. Moreover, the findings from this study offer advancements in annotation efficiency and present a robust tool for body composition analysis, with potential applications in enhancing diagnostic and prognostic assessments in clinical settings.