BodyMAE: A Surface Area-Aware Masked Autoencoder for Body Composition Estimation from 3D Body Scans

Accurate assessment of body composition is important for risk stratification and management of metabolic, musculoskeletal, and aging-related diseases, yet reference modalities such as dual-energy X-ray absorptiometry (DXA) may have limited accessibility for frequent monitoring. Commodity 3D body scans offer a radiation-free alternative, but extracting predictive shape features remains challenging because of substantial variation in body size, morphology, and acquisition conditions. Therefore, we introduce BodyMAE, a self-supervised, surface area-aware masked autoencoder for body composition estimation from metric-scale 3D body scans. The framework combines area-adjusted sampling, long-range feature learning, and Repulsion Loss to promote locally uniform reconstruction. Evaluated on 917 paired 3D body scans and DXA measurements, BodyMAE achieves RMSEs of 3.92 percentage points, 3.75 kg, and 3.39 kg for fat percentage, fat mass, and lean mass, with corresponding \(R^{2}\) values of 0.91, 0.97, and 0.91. BodyMAE also improves representation consistency across repeated scans, achieving an mAP of 86.7%. These results demonstrate the effectiveness of BodyMAE for self-supervised representation learning and body composition assessment from 3D body scans.