3D convolutional deep learning for nonlinear estimation of body composition from whole body morphology

Area of Science:

• Biomedical Engineering
• Computer Vision
• Human Physiology

Background:

• Previous body composition prediction from 3D optical imagery relied on linear algorithms.
• A need exists for more accurate and precise methods in human body shape parameterization and composition estimation.

Purpose of the Study:

• To apply deep 3D convolutional graph networks and nonlinear Gaussian process regression (GPR) for body composition estimation.
• To compare the performance of nonlinear GPR and deep shape features against linear models.

Main Methods:

• Utilized a novel ensemble body shape dataset with 4286 3D scans.
• Developed and tested deep 3D convolutional graph networks and nonlinear Gaussian process regression.
• Performed ablation studies comparing linear and nonlinear models.

Main Results:

• Nonlinear GPR reduced prediction error by up to 20% and increased precision by 30% compared to linear regression.
• Deep shape features improved prediction error by 6-8% for males and precision error by 4-14% for both sexes.
• All models achieved coefficients of determination (R² > 0.86) and lower estimation RMSEs than prior studies.

Conclusions:

• Deep learning and nonlinear GPR offer superior performance for body composition prediction from 3D imagery.
• These advanced methods provide more accurate and precise estimations of body composition metrics.
• The findings advance the field of non-invasive body composition analysis.