Kimberly Gerard
Research Mentor: Sujata Khandare
Mentor Department: University of Michigan Transportation Research Institute, Engineering
Author(s): Kimberly Gerard, Sujata Khandare
Session: Session 2 (10:00 AM – 10:50 AM)
Presentation Type: Poster 79
Abstract
Motor vehicle crashes are a leading cause of musculoskeletal injury, but the mechanisms that drive these injuries, and how they vary across individuals, are not fully understood. Human body models provide a powerful way to investigate these mechanisms and simulate tissue loading under crash conditions. However, the fidelity of these simulations depends on the accuracy of the underlying anatomy of the models. Precise three-dimensional segmentations from medical imaging are needed to capture subject-specific geometry, enable consistent mesh/model generation and support realistic boundary conditions. Segmentation errors can propagate directly into model geometry and can compromise predicted kinematics, load paths, and injury metrics. Therefore, high-quality segmentation is a critical prerequisite for building reliable human body models and enabling population-scale studies of injury risk. This study used TotalSegmentator, a deep learning-based tool for automated anatomical segmentation from CT scans, implemented within 3D Slicer, to segment upper-extremity skeletal anatomy from CT data collected at Michigan Medicine. Upper-extremity CT scans were obtained from 328 adult participants under an Institutional Review Board (IRB)-approved protocol. The cohort was approximately sex-balanced and spanned a wide range of age (18–79 years), stature (1.4 – 1.8 m), and BMI (15.9–53.5 kg/m²) to capture anatomical variability and support evaluation across a diverse set of individuals. To enable efficient processing at scale, a custom batch-processing script was developed to automate segmentation across all scans. The resulting segmentations were then manually reviewed and quality-checked for anatomical consistency, including verification of correct structure identification and detection of mislabeling or segmentation errors requiring correction. Following quality control, the verified bone geometries, particularly the humerus, radius, and ulna, will be used to develop a statistical shape model and will ultimately be integrated into a human body modeling framework to support analyses of musculoskeletal injury patterns associated with motor vehicle crashes. This workflow enables scalable, population-level model development across a diverse cohort, supporting investigation of how individual characteristics, such as age, stature, sex, and BMI, may contribute to differences in anatomy and, ultimately, injury risk under crash loading. 328 CT scans were successfully segmented, and manual quality control was completed for each case to confirm anatomical consistency. Any mislabeling or segmentation errors identified during review were corrected to produce a verified dataset of upper-extremity bone geometries. Future work will expand the framework by incorporating musculature into the three-dimensional model to further improve the biofidelity of the model. Beyond crash injury biomechanics, this pipeline has potential applications in preoperative surgical planning and patient-specific modeling.
