Learning Clinical Body Composition Metrics from 2D and 3D Optical Imaging

Accurate human body shape representation has many applications within computer graphics including 3D animation, virtual tailoring, ergonomic engineering, and virtual reality reconstruction. For clinical researchers working at the intersection of computer graphics, machine learning, and obesity-related epidemiology, computational modeling of human body shape presents a novel and accessible pathway to quantifying, classifying, and monitoring risk factors associated with premature mortality caused by metabolic syndrome. Total and regional body composition and body shape are strongly correlated with progression of metabolic syndrome as well as degenerative conditions such as sarcopenia and osteoporosis. Estimates of body composition from optically measured human body geometry are cheap, safe, and non-invasive relative to current reference methods that require exposure to ionizing radiation.In this thesis, I present a body of work that thoroughly investigates predicting body composition from optical images of human body shape with clinically significant precision and accuracy. This thesis contributes the following: 1. Introduces a model that predicts 3D body shape and total and regional body composition metrics from monocular 2D images. 2. Develops a method that automatically standardizes 3D human body scans to watertight manifold mesh templates with consistent topology and anatomical correspondence and demonstrates the viability of this tool for constructing new shape and regression models that predict body composition with agnosticism towards input scanning devices. 3. Extends the automatic mesh templating method to create the first autoencoded shape model for a pediatric cohort paired with body composition prediction from shape parameters derived with unsupervised learning. 4. Performs a systematic review of deep 3D shape autoencoders for total human body geometry with the goal of identifying the current state of the art methods and architectures in reconstruction accuracy while also suggesting standards and best practices for future work in this research field. 5. Performs a novel estimation of body composition from nonlinear features extracted by a deep autoencoder with nonlinear Gaussian process regression and comprehensively compares marginal contributions of linear and nonlinear shape and regression algorithms against the linear baselines of prior works with systematic ablation studies.

Author dc:creator

• Tian, Isaac Yuheng

Advisor dc:contributor.advisor

• Curless, Brian

• 3D deep learning
• 3D Optical Scanning
• Body composition prediction
• Human body shape
• Medical physics
• Statistical shape model
• Computer science
• Nutrition
• Medical imaging

Statement dc:rights

• none

Language dc:language.iso

en_US

Handle dc:identifier.uri

http://hdl.handle.net/1773/50759

OAI identifier oai:identifier

oai:digital.lib.washington.edu:1773/50759