Deep learning for opportunistic screening for osteoporosis and osteopenia from radiographs - PROJECT SUMMARY/ABSTRACT Low bone mineral density (BMD) is an innocuous and pernicious disease that eventually results in osteopenia and osteoporosis, and increased risk for osteoporotic fractures. These fractures result in increased morbidity due to pain, decreased locomotion, loss of income, and increased healthcare cost. Hip fractures have been associated with increased mortality, with up to 35% 3-year mortality rate. Unfortunately, most patients at risk for fractures are not screened for low BMD using dual energy x-ray absorptiometry (DXA). There is a need to identify patients with low BMD to institute therapy/treatment to prevent the development of fractures. Approximately 280 million radiographs are performed annually in the United States. Our central hypothesis is that deep learning analysis of these radiographs could provide information about the underlying patient bone quality and predict whether a patient is osteoporotic or osteopenic. The long-term goals are (1) to identify imaging biomarkers (radiomic features) that are associated with osteoporosis and osteopenia and (2) to identify imaging biomarkers that predict underlying trabecular bone scores (TBS). This proposal utilizes cutting edge computational and statistical approaches based on radiographs of patients with concurrent DXA studies. The central hypothesis will be tested by pursuing three specific aims: (1) determine key reproducible imaging biomarkers (radiomic features) of bones from radiographs in patients without implanted metallic devices that could be used to reliably predict whether a patient has osteoporosis or osteopenia; (2) identify imaging biomarkers (radiomic features) of bones from radiographs in patients with implanted metallic devices that could be used to reliably predict whether a patient has osteoporosis or osteopenia; (3) determine whether these imaging biomarkers from radiographs could be used to predict the bone TBS. These aims will be pursued using novel computational and statistical techniques from the imaging sciences. We will identify imaging features from radiographs that can predict whether a patient has osteoporosis or osteopenia and who should go on to further screening with DXA, which would be a significant advancement for clinical healthcare. We also aim to identify imaging features that can predict TBS, which provides additional information about the bone microarchitecture and future fracture risk. The impact of this work is that we will have a better understanding of the manifestations of osteoporosis and osteopenia from radiographs. The results of this work will identify patients who have undergone radiographs who should go on to have formal DXA evaluation because of their increased risk for osteoporosis and osteopenia.
