SHAPEDNET: a regression convolutional neural network approach for body composition assessment based oninterpretable DBF-SCAN techniques

• Ano de defesa: 2023
• Autor(a) principal: Nascimento, Navar de Medeiros Mendonça e
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Palavras-chave em Português: Composição Corporal - Avaliação; Distribuição da gordura corporal; Absorciometria de fóton; Aprendizado multitarefa; Aprendizagem profunda; Inteligência artificial; Body composition evaluation; Body fat distribution; Absorptiometry, Photon; Multi-task learning; Deep learning; Artificial intelligence; CNPQ::ENGENHARIAS::ENGENHARIA ELETRICA
• Link de acesso: http://repositorio.ufc.br/handle/riufc/79301

Accurate body composition estimation is essential for assessing health and well-being. This thesis introduces a novel techniques for improving body composition assessment, the ShapedNet. This method is a multi-task learning deep neural network that simultaneously estimates Body Fat Percentage (BFP), identifies, and localizes individuals using a single photo. To validate its accuracy, rigorous comparisons were made against the gold standard reference method, Dual-Energy X-ray Absorptiometry (DXA). The validation process involved a dataset of 1273 healthy adults, ranging in ages (18-65 years old), sexes (54.59% women), and BFP levels (DXA 9.3%- 57.6%). The evaluation included BFP measurements using DXA, two photo sections, and other clinical exams. Additionally, a follow-up dataset was created by repeating all exams and photos with the participants three months later. Our approach outperformed the best reported results for computer vision-based body fat estimation using photos by a margin of 1.19 to 1.89 points. ShapedNet achieved a Mean Absolute Percentage Error (MAPE) of 4.91% and Mean Absolute Error (MAE) of 1.42. We also evaluated using both gender-based and non-gender-based approaches, the later outperformed the gender-based by a margin of 1.07 to 2.12 points. ShapedNet estimates BFP with 95% confidence, within an error margin of 4.01% to 5.81%, or a difference of 1.18 to 1.66 in DXA estimates. High correlation coefficients greater than 0.9 were observed between ShapedNet and DXA in all experiments conducted. To enhance the interpretability of the proposed method, we design the DBF-SCAN visualization technique. DBF-SCAN provides valuable visual insights into the distribution and composition of body fat, highlighting the relative importance of different body regions in determining body fatness. This subject-oriented visualization complements the quantitative estimation provided by ShapedNet. This research introduces a robust and accurate method for body composition assessment, supported by rigorous validation, correlation analysis, and visualization techniques. The ShapedNet method aligns with the two-compartment method in body composition assessment theory and offers an innovative indirect approach. Beyond body composition estimation, the advancements presented in this thesis contribute to the development of Explainable AI (XAI) models, and broaden the applications of multi-task learning in various domains.