New research presents a novel method for estimating body composition using 3D imaging and deep learning. Traditional methods like DXA scans, while accurate, involve radiation exposure. This research explores a non-invasive optical approach as a potential alternative.
The study utilizes a dataset of 4286 3D body scans, some paired with DXA scan data for ground truth. Researchers developed a deep 3D graph convolutional autoencoder (3DAE) to extract shape features from the 3D scans. They then employed nonlinear Gaussian process regression (GPR) to predict body composition variables from these features. The performance of this nonlinear approach (3DAE-GPR) was compared against traditional linear methods like principal component analysis (PCA) combined with ordinary least squares regression (OLS).
Results showed that the nonlinear GPR significantly improved prediction accuracy and precision compared to linear regression, achieving up to a 20% reduction in prediction error and a 30% increase in precision for various body composition metrics. While deep shape features from the 3DAE improved predictions for males, they did not consistently outperform linear PCA features for females. However, the 3DAE did improve precision for both sexes. Critically, the 3DAE-GPR model achieved lower estimation errors than any previously reported method for ten key body composition variables, demonstrating the potential of this approach for accurate and non-invasive body composition analysis. The research also explored the impact of pre-training the 3DAE on a large dataset of human poses (DFAUST) and the importance of including diverse datasets for training. The study suggests that combining 3D optical imaging with deep learning and nonlinear regression offers a promising pathway for improved body composition assessment, potentially impacting the management of various health conditions.
Related Content: PSMA-PET Imaging For Prostate Cancer Detection