Quantifying body composition from computed tomography (CT) provides valuable insights into metabolic health, disease prognosis, and treatment outcomes. However, the development of 3D segmentation models for body composition analysis has been limited by the extensive manual annotation effort required. We present an annotation-efficient strategy for 3D segmentation of abdominal and pelvic body composition [1], designed to drastically reduce annotation needs while maintaining high accuracy. Our approach combines sparse manual annotations with an iterative self-learning framework that transitions from 2D to 3D segmentation. Only 1% of all training slices were manually annotated. The model was trained on 116 CT scans and evaluated on an internal test set of 20 scans and a reader study of 100 cases. Quantitative performance was assessed using the Dice similarity coefficient. To further assess generalizability and clinical reliability, a multi-reader evaluation was conducted by three experienced radiologists using a standardized scoring protocol to rate the correction effort per segmentation class. The final 3D model achieved Dice coefficients of 0.97 ± 0.01 for skeletal muscle (SM), 0.85 ± 0.04 for inter-/intramuscular adipose tissue (IMAT), 0.94 ± 0.04 for visceral adipose tissue (VAT), and 0.98 ± 0.01 for subcutaneous adipose tissue (SAT). Reader study results confirmed negligible to minimal correction effort for SM, VAT, and SAT, with higher variability for IMAT. These results indicate strong robustness and demonstrate the feasibility of developing accurate 3D body composition models with minimal annotation effort.