Osteoporosis is the pathological disorder of bones, characterised by decreased bone mineral density (BMD) and microarchitectural deterioration of bone tissue, leading to increased fracture risk. Early observation and treatment of fracture risk from potential patients can reduce the incidence and medical expenses. While much of clinical fracture risk assessment is carried out through dual-energy X-ray absorptiometry (DXA) imaging, high-resolution peripheral quantitative computed tomography (HRpQCT) is becoming increasingly available, providing more detailed analysis of bone microarchitecture at the peripheral skeleton. Traditional analysis of HR-pQCT images requires manual operation and results in a multitude of cortical and trabecular parameters which would be potentially cumbersome to interpret for clinicians. Automated quantitative analysis of HR-pQCT scans to ascertain fracture risk, would be far simpler and more effcient.

Our research work primarily focuses on the dataset from the Hertfordshire Cohort Study (HCS), which comprises 2997 men and women born in Hertfordshire from 1931- 1939 and who still lived there in 1998-2004. 376 participants of the HCS attended research visits at which clinical covariates were measured;​ fracture history was determined via self-report and vertebral fracture assessment. Bone microarchitecture was assessed via HR-pQCT scans of the non-dominant distal tibia, and BMD measurement and lateral vertebral assessment were performed using DXA. In addition, our study utilises the dataset from the Global Longitudinal Study of Osteoporosis in Women (GLOW) which involves 723 physicians and 60,393 women aged 55 years and older in 10 countries. In this cohort, 501 participants completed self-administered questionnaires and underwent HR-pQCT scans of the non-dominant distal radius and tibia, as well as DXA scans of whole body, proximal femur and lumbar spine.

Building upon the correlation between bone microarchitecture and fracture risk, we develop an automatic approach to discriminate previous fractures by using HR-pQCT measures of bone microarchitecture. We propose a method based on local binary pattern (LBP) to characterise the texture patterns of HR-pQCT images and to quantify bone microarchitecture via statistical distributions. Further, we decouple the relative contributions of cortical and trabecular compartments in fracture discrimination. Our method includes a deep neural network-based segmentation algorithm for separating the cortical and trabecular regions to enable texture features to be extracted separately and their statistical distributions quantifed.

In addition to volumetric texture analysis, we present a novel discriminative system to automatically identify individuals with previous fractures from HR-pQCT images using a combination of multi-view convolutional neural networks (CNNs) and the random forest algorithm. Unlike conventional deep learning architectures that require a massive amount of training data, our method based on transfer learning extracts image features from representative views of HR-pQCT scans to characterise bone microarchitecture and then integrate the features for fracture discrimination.

Last but not least, we propose an adaptive threshold strategy to further enhance the accuracy and robustness of our discriminative system for previous fracture. Our method generates adaptive thresholds based on DXA-measured T-scores of the participants within the population to flter out healthy subjects with traumatic fractures and osteoporotic non-fractured subjects. Then we adopt multi-view CNNs to characterise bone microarchitecture in HR-pQCT images to distinguish between non-fractured healthy subjects and subjects with osteoporotic fractures. Furthermore, we evaluate the performance of our discriminative system on an independent cohort.