Abstract
Decline in paraspinal muscle health, characterized by increased intramuscular fat (IMF) and reduced muscle size, has received increasing attention as a feature of spinal health and a contributor to the development and recurrence of low back pain (LBP). However, the relationship between paraspinal muscle health decline and LBP is complex and influenced by various mechanical and metabolic factors. Clinical and methodological variability, along with the time-consuming and labour-intensive nature of assessing paraspinal muscle health from magnetic resonance imaging (MRI), further complicates our fundamental understanding of why, where, and to what extent muscles change resulting in, or following, LBP. Consequently, the clinical meaning of paraspinal muscle health decline in individuals with LBP remains inconclusive. Therefore, the overall aim of this thesis was to a) develop and evaluate the performance of computer vision models for the automated, standardized, and detailed quantification of paraspinal muscle health from MRI, and b) investigate the clinical meaning of paraspinal muscle health decline in individuals with LBP. Chapter 2 evaluates the performance of Convolutional Neural Networks (CNN) for the automatic segmentation of the paraspinal muscles from MRI. This study highlights that CNNs are highly efficient and accurate in segmenting the paraspinal muscles, achieving human-level performance in only a fraction of the time (human ~35 minutes versus CNN ~6 seconds per subject). Furthermore, the findings support the use of 2D models over 3D models and showed that training without data augmentation performed better than training with data augmentation. Chapter 3 evaluates the performance of automated thresholding techniques for the quantification of lumbar paraspinal IMF. In this study, we found that the performance of automated thresholding techniques is dependent on the choices of algorithms, the number of components, and the muscles assessed. For the lumbar multifidus and erector spinae, Gaussian Mixture Modelling initialized with 3 components performed best, while none of the algorithms performed well for the psoas major. Chapter 4 evaluates the feasibility of template-based spatial parametric mapping of paraspinal IMF. In this study, we developed a spinal muscle template and demonstrated its feasibility in 76 participants who had recovered from LBP. Preliminary spatial parametric maps indicated that the associations between paraspinal IMF and age, sex, and BMI are localized differently both along the superior-inferior expanse of the lumbar spine and spatially in the transverse plane. Chapter 5 describes the association between paraspinal muscle health and age, sex, body mass index (BMI), physical activity, and LBP in 9,564 participants from the UK Biobank using a computer vision model. In this study, we found that paraspinal muscle health decline was associated with age, sex, BMI, and physical activity. The results underscore that a decline in paraspinal muscle health is not exclusively indicative of structural pathology in individuals with LBP. Nevertheless, this study found that people with LBP have higher levels of IMF and lower muscle size compared to those without LBP. Additionally, we found that these differences were not localized to the lower lumbar region but were more broadly distributed across the lumbar spine. Chapter 6 describes the association between pre-operative paraspinal IMF and 5-year clinical recovery following lumbar surgery for spinal stenosis. In this study, we found that higher levels of IMF in the multifidus, but not in the erector spinae, were associated with an unfavourable clinical course. Chapter 7 describes a systematic review of the reversibility of paraspinal muscle IMF in people with LBP through exercise. In this study, we found moderate-quality evidence that paraspinal IMF is not reversible through exercise. Inconsistency in image acquisition and IMF quantification methods and variations in exercise descriptions raise concerns about the generalizability and applicability of the treatment effect estimates included.
Original language | English |
---|---|
Qualification | PhD |
Awarding Institution |
|
Supervisors/Advisors |
|
Award date | 8 Apr 2025 |
Print ISBNs | 9789465069227 |
DOIs | |
Publication status | Published - 8 Apr 2025 |