DXA-based statistical shape-intensity models for hip fracture prediction in post-menopausal women

Osteoporosis is a bone disease caused by an imbalance between bone deposition and resorption which affects bone microstructure and leads to a decreased bone mass. According to recent studies, over 200 million people worldwide have osteoporosis. There are approximately 9 million fractures worldwide per year due to osteoporosis, among which 1.6 million are hip fractures (75% of which affect women). Moreover, it is estimated that 1 in 3 females and 1 in 5 males over the age of 50 will have an osteoporotic fracture. Osteoporosis is particularly frequent in post-menopausal women for hormonal reasons. The current gold standard to diagnose osteoporosis is measuring the areal Bone Mineral Density (aBMD) through dual-energy X-ray absorptiometry (DXA); aBMD is then used to calculate the so-called T-score, an indicator based on the comparison between the screened subject’s BMD and the BMD of young females aged in the range of 20-29 years. The World Health Organization (WHO) has chosen T-score as the ultimate parameter to discriminate osteoporotic subjects from non-osteoporotic ones; however, its performances in the fracture prediction field are limited: it is estimated that about one half of the subjects not classified as osteoporotic by the T-score did in fact fracture. Several alternatives to T-score have been investigated, like epidemiological models, structural parameters and FE models, all with their own limitations. Thus, more attention has been given to the so-called statistical models. This work was based on the development the afore-mentioned statistical models starting from DXA images, which would be already clinically available. The aim of the work was to assess if they could support an enhanced identification of the risk of hip fracture with respect to the gold standard T-score. A retrospective cohort of 97 British post-menopausal women was herein considered. For these subjects, both CT and DXA images were available. Out of the 97 patients, 49 subjects had experienced a femoral fracture, while 48 had not. For each subject, the DXA image was segmented to extract the proximal femur’s shapes. Hence, Statistical Shape Models (SSMs) based on Principal Component Analysis (PCA) and Partial Least Square (PLS) could be built. PCA aims to find new variables that maximize the feature’s variance, while PLS maximizes the covariance with an external variable (which, in this work, is the fracture status). Furthermore, also PCA- and PLS-based Statistical Intensity Models (SIMs) were built, taking advantage of the available local aBMD maps. Then, the outcomes of the two kinds of model were combined and further processed through PCA and PLS to create two Statistical Shape-Intensity Models (SSIMs). The calculated PLS modes highlight that neck length, size and inclination are the shape features associated with fracture, while as for intensity a generally decreased density throughout the femur and a decreased cortical thickness are the main features associated with fracture. The logistic regression models built using the whole cohort as training set show that, in terms of Receiver Operating Characteristic (ROC) curves and Area Under Curve (AUC), the PLS-based SSIM show classification performances (AUC ≈ 0.85) improved with respect to aBMD (AUC ≈ 0.74). Eventually, the built statistical models were also validated through a 10-fold cross validation process able to assess the predictive performances of the models in separating fractured from non-fractured cases.