Creating representative skeletal cohorts through statistical modelling and synthetic CT image generation
| dc.contributor.author | Beagley, Aren | |
| dc.contributor.supervisor | Giles, Joshua W. | |
| dc.date.accessioned | 2026-01-08T21:50:05Z | |
| dc.date.available | 2026-01-08T21:50:05Z | |
| dc.date.issued | 2025 | |
| dc.degree.department | Department of Mechanical Engineering | |
| dc.degree.level | Master of Applied Science MASc | |
| dc.description.abstract | Finite Element (FE) analysis is an important tool for orthopaedic research that allows studying the effects of orthopaedic devices, such as joint replacements, in ways that would be difficult or impossible to investigate experimentally. High fidelity FE studies are typically performed using a cohort of subject-specific FE bone models created from Computed Tomography (CT) images to ensure realistic bone shapes and material properties. However, this limits sample sizes because acquiring CT images exposes subjects to harmful radiation and such exposure should be avoided unless medically necessary. Unfortunately, small sample sizes create significant limitations on the applicability and generalizability of the insights provided by most FE studies. Ideally, FE studies should use sufficiently large and diverse cohorts to ensure that the resulting insights generalize to the broader population. If the population-level distribution of shape, size, and stiffness for a bone was known prior to conducting the FE study, subjects could be systematically chosen to create a representative cohort; however, even the process of choosing these subjects can be a challenge. Statistical Shape and Intensity Models (SSIMs) are an established tool for characterizing the population-level variance of size, shape, and material properties, and these models can be created from pre-existing, medically necessary, CT images. Furthermore, SSIMs can generate new instances that are representative of the population. Unfortunately, previous methods of developing SSIMs do not capture sufficient detail and produce models incapable of generating new instances suitable for use in FE studies. This thesis describes the development and validation of a method for creating high resolution SSIMs capable of generating new instances that contain data comparable to that of CT images. An associated method for converting SSIM-generated instances into SSIM-derived synthetic images, that are comparable to CT images, was also developed and validated. In combination, these two methods result in a model capable of systematically sampling a population-level distribution to create representative cohorts suitable for use in FE studies. To determine how representative the combined method is, generalization of both SSIM-generated instances and SSIM-derived synthetic images were assessed by comparing shape and material properties against real subjects. SSIM-generated instances were assessed to have Mean Generalization Root-Mean-Square (MGRMS) errors of 2.15 mm and 228.1 Hounsfield Units (HU) for shape and material properties respectively, and an average surface distance (ASD) of 1.145 mm. After converting SSIM-generated instances to synthetic images, the MGRMS error for shape could not be assessed, but for material properties it increased to 286.2 HU (+58.1) and the ASD showed improvement by decreasing to 1.028 mm (-0.117). These results demonstrate that the high resolution SSIM presented in this work has similar generalizability as previously published SSIMs while capturing significantly greater variance and that the conversion to synthetic images introduces minimal additional error. As such, the combined high resolution SSIM and conversion algorithm developed for this thesis represent a viable method of improving generalization for FE studies by overcoming the current barriers to systematically producing representative cohorts. This work set the stage for future work investigating the impacts of conducting FE studies with SSIM-derived synthetic images and addressing a range of clinically relevant biomechanical questions. | |
| dc.description.scholarlevel | Graduate | |
| dc.identifier.uri | https://hdl.handle.net/1828/23059 | |
| dc.language | English | eng |
| dc.language.iso | en | |
| dc.rights | Available to the World Wide Web | |
| dc.subject | Statistical Shape and Intensity Model | |
| dc.subject | Partial Volume Artifact | |
| dc.subject | Non-Rigid Registration | |
| dc.subject | Principal Component Analysis | |
| dc.subject | Synthetic Computed Tomography | |
| dc.title | Creating representative skeletal cohorts through statistical modelling and synthetic CT image generation | |
| dc.type | Thesis |