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

Area of Science:

• Biomedical Engineering
• Computer Vision
• Medical Imaging

Background:

• Accurate body composition assessment is vital for managing chronic diseases.
• Current gold-standard methods like DXA are expensive and not practical for frequent use.
• 3D body scans present a cost-effective, radiation-free alternative, but extracting useful data is challenging.

Purpose of the Study:

• To develop a novel method, BodyMAE, for accurate body composition estimation from 3D body scans.
• To overcome challenges in processing 3D scans, including variable density and scale.
• To validate BodyMAE's performance against Dual-energy X-ray absorptiometry (DXA) measurements.

Main Methods:

• Developed BodyMAE, a self-supervised masked autoencoder tailored for metric-scale 3D body scans.
• Integrated surface-area aware sampling and a long-range focused encoder.
• Trained and evaluated the model on 917 paired 3D body scans and DXA reports.

Main Results:

• BodyMAE achieved high accuracy in estimating fat percentage (RMSE 3.825), fat mass (RMSE 3.694 kg), and lean mass (RMSE 3.608 kg).
• Demonstrated competitive performance for bone mineral content estimation (RMSE 0.284 kg).
• BodyMAE's learned representations showed superior feature stability and retrieval accuracy compared to baselines.

Conclusions:

• BodyMAE effectively enables accurate body composition estimation from 3D body scans.
• The combination of metric-aware sampling, relational encoding, and geometric regularization is key to the model's success.
• This approach provides a viable, low-cost alternative for frequent body composition monitoring.