Body Composition Estimation from 3-Dimensional Optical Scans Using Principal Component Analysis in Youth Soccer Players

This thesis aims to predict total and regional body composition derived from dual-energy X-ray absorptiometry (DXA) using principal component analysis (PCA) on raw and digitally reposed three-dimensional optical (3DO) whole-body scans in youth soccer players. This work is motivated by the growing demand for safe, accessible, and low cost methods of assessing body composition, since the current gold standard, DXA, is often impractical due to cost, exposure to ionizing radiation, and limited accessibility, particularly in youth sports. The purpose of this study is to evaluate the feasibility and accuracy of this approach to determine whether it can serve as a valid and reliable tool for monitoring body composition. A total of 429 participants, including 270 males and 159 females between 14 and 23 years old, took part in the study. The acquisition campaign included body weight and height measurements, one whole-body DXA scan, and two 3DO scans acquired via Mobile Fit app. Two different analyses were carried out on the avatars: A-pose and T-pose. The raw A-pose avatars were used as provided by the Mobile Fit app, without any processing. The T-pose avatars, on the other hand, were sent to Meshcapade to be digitally reposed to a standardized pose. PCA was applied to perform dimensionality reduction and capture complex shape features. The algorithm was independently applied to male and female avatars. The datasets, which included principal components (PCs), anthropometric variables, and DXA outcomes, were randomly split into training (80%) and test (20%) sets. Stepwise linear regression was used to construct prediction models for DXA body composition outcomes. The performance of the models was evaluated using fivefold cross-validation applied to the training set. The results showed that estimates of total and regional lean mass were highly accurate in male soccer players, with coefficients of determination (R2) above 0.82 for hybrid models that combine PCs and digital anthropometry, and up to 0.92 for total lean mass. In contrast, fat mass estimates were inaccurate since they reported R2 values close to or below 0. In female soccer players, none of the models achieved high performance in total and regional lean mass estimates (R2 as high as 0.62 for total lean mass), and fat mass predictions achieved moderate accuracy (R2 as high as 0.4 for body fat percentage). This study demonstrated that accurate predictions were reached for specific outcomes and subgroups, suggesting that the homogeneity in body shape and composition among young athletes may limit the ability of PCA to capture body shape variance. In this context, models that relied only on digital anthropometric measurements performed as well as, or better, than those that used PCs.