Novel imaging and processing for early musculoskeletal disease

Musculoskeletal disease, especially osteoarthritis (OA) is a major cause of pain and disability. The combination of large burden and limited treatment options for patients leads to a need to improve our understanding of how it manifests and progresses in its early stages, and to link this understanding to individual patients. Quantitative magnetic resonance imaging (MRI) techniques has demonstrated the ability to detect these early changes, among which T2 mapping is one of the most widely used. This thesis attempts to translate and build a platform technology for the analysis of musculoskeletal MRI, with a focus on T2 mapping. By performing a systematic review and meta-analysis, the results of 12 knee studies, and 4 knee studies are combined, from which the difference in T2 value between OA and healthy tissue are found to be extremely small (ΔT2META (Knee) = 2.20 ms, ΔT2META (Hip) = 3.31 ms). Using these values as the bottom line criteria and by performing a Monte Carlo simulation, the performance of 16 T2 fitting methods are further evaluated. It is found although advanced T2 fitting methods such as NCEXP (where a simplified equation is reported for the first time) and SQEXP can achieve good trueness performance, the poor precision caused by noise would make absolute measurement of T2 of little use. A new quantitative approach, which involves transforming T2 images from different patients to the same space is needed to allow direct comparison and better analysis. OxteoFlow is the foundation of the analysis platform developed in this thesis, which uses affine transformations followed by a more complex nonlinear transformation to register individuals to a reference scan. As can be seen by visual inspection, OxteoFlow successfully normalised knees and hips of different shapes into their respective reference spaces. Parallel computing and grid search optimisation are enabled in OxteoFlow, making it a suitable tool for musculoskeletal imaging studies. Using data from the Osteoarthritis Initiative (OAI), the baseline T2 scans are stratified into four groups—‘Both’, ‘ROA’, ‘Pain’ and ‘None’ based on their outcomes. By performing a voxel-based pattern analysis on their baseline T2 maps, 6 different voxel activation patterns, all relating to different incident/progression outcomes are identified, with their complex relationship illustrated by voxels detected distance maps. Pattern 6, which shows effect of ROA Progression activations is found to be the most significant as it survived the extremely conservative Bonferroni correction at P < 0.001, indicating T2 mapping is a powerful tool to predict ROA progression outcome at the earliest possible time. Specifically targeting the ROA progression cohort, a deep learning model, OxteoNet, is developed and compared against leading machine learning algorithms. It is found the OxteoNet provides a small, though significant, improvement in AUROC, yet a dramatic improvement in AP score than traditional machine learning methods. The class activation maps (CAMs) of OxteoNet may serve as an indicator for the clinician as sites at which pathology are observed that need further inspection. In summary, this thesis covered a complete research and development (R&D;) life cycle of T2 imaging and its applications (calculation → preprocessing → analysis → diagnosis/prognosis), answered some fundamental and important questions related to this non-invasive OA biomarker, translated some of the latest and best technical advances from a wide range of medical imaging domains into musculoskeletal, and demonstrated its value as an image analysis platform technology for musculoskeletal studies.