3D Convolutional Deep Learning for Nonlinear Estimation of Body Composition from Whole-Body Morphology.

Area of Science:

• Biomedical Engineering
• Medical Imaging
• Computer Science

Background:

• Accurate body composition estimation is crucial for metabolic syndrome assessment.
• Prior non-invasive methods using 3D scans and linear models had limitations in precision and accuracy.
• Dual X-ray absorptiometry (DXA) is the reference standard for body composition measurement.

Approach:

• Developed a novel nonlinear approach using deep 3D convolutional graph networks for body composition modeling.
• Trained a parameterized shape model with a graph convolutional 3D autoencoder (3DAE) on an ensemble dataset of 4286 3D scans.
• Utilized nonlinear Gaussian Process Regression (GPR) on 3DAE features to predict body composition against DXA measurements.

Key Points:

• Nonlinear GPR demonstrated up to 20% reduction in prediction error and 30% increase in precision over linear regression.
• Deep shape features improved prediction accuracy for males and precision for both sexes.
• The best nonlinear model outperformed prior linear methods on all body composition prediction metrics, with R² > 0.86.

Conclusions:

• Gaussian Process Regression (GPR) offers superior precision and accuracy for body composition mapping from shape features compared to linear regression.
• Deep 3D features enhance body composition prediction precision in both sexes and accuracy in males.
• This nonlinear deep learning approach achieves lower Root Mean Square Errors (RMSEs) than previous methods for 10 body composition metrics.