VLBODY — Vision-learned body composition estimator: uma abordagem profunda para análise visual da composição corporal

Abstract

Accurate assessment of body composition is essential for monitoring nutritional status, physical performance, and metabolic risk in individuals. Traditional methods, such as dual-energy X-ray absorptiometry (DXA), although highly accurate, present limitations related to cost, accessibility, and large-scale applicability. In this context, this thesis proposes VLBODY — Vision-Learned Body Composition Estimator, a deep visual learning methodology designed for the noninvasive estimation of body fat percentage from two-dimensional images. The proposed method is organized as a sequential pipeline comprising three main stages: (A) detection, responsible for the automatic identification of the region of interest (ROI) corresponding to the human body in the image; (B) segmentation, aimed at the precise isolation of the body silhouette and removal of background noise; and (C) body fat estimation, which relies on the visual encoding component of Vision–Language Model (VLM) architectures to extract morphological representations from the ROI and, through a regression process, convert them into continuous body fat percentage values. This unified architecture integrates the entire process, from image input to final prediction, ensuring coherence between feature extraction and quantitative inference. Four architectural variants (VLBODY-S, VLBODY-M, VLBODY-L, and VLBODY-X) were implemented, with different depths and latent dimensions, to investigate the impact of representational capacity on predictive performance. The method was validated on the BCDB23 dataset, composed of 1,044 participants whose standardized images were compared against DXA reference measurements. The VLBODY-S and VLBODY-X variants achieved the best results, reaching coefficients of determination (R²) above 0.8 and mean absolute errors (MAE) below 3 percentage points, outperforming traditional anthropometric-based methods and approaching state-of-the-art techniques. In addition to quantitative performance, VLBODY incorporates an interpretability mechanism based on spatial relevance maps, highlighting the body regions that most influence predictions. These results indicate that VLBODY represents a promising alternative for noninvasive body composition assessment, combining accuracy, interpretability, and potential applicability in both clinical and population-based contexts.