Predictive modelling of femur fracture from DXA scans using radiomics and machine learning

Abstract

Predictive modeling of bone fractures at the hip using DXA images is an important research field due to the capability of this imaging modality to capture osteoporosis related changes in the bone tissues. There exist many available techniques providing fracture risk assessment, such as bone mineral density measurements and biomechanical models. However, these methods do not use the wealth of information provided by the DXA scans and thus lack the accuracy to enable their translation to clinical practice. In this study, a radiomics and machine learning approach is proposed for a more comprehensive predictive modelling of femur fracture using DXA. Our main hypothesis is that integrating heterogeneous and complex characteristics of the bone tissue through radiomics at both the global and local scales will lead to improved prediction of fracture risk. In the proposed technique, the optimal radiomics indices of different types (shape, intensity and texture based) are selected using feature selection methods to identify the most relevant ones for discriminating low-risk and high-risk cases. Furthermore, advanced machine learning is applied to integrate the selected radiomics features into a unified risk classification model based on different learning models (Support Vector Machines, Decision Tree and Random Forest). The proposed predictive model was validated using 63 cases including to patients with and without femur fracture. In this preliminary study, all cases were correctly classified using the proposed model, indicating great potential of radiomics-based classification for predicting fractures of the femur.