Hadi Salehpour - Student Research Committee, Tabriz University of Medical Sciences, Tabriz, Iran
Morteza Ghojazadeh - Student Research Committee, Tabriz University of Medical Sciences, Tabriz, Iran
Alireza Lotfi - Student Research Committee, Tabriz University of Medical Sciences, Tabriz, Iran
Ali Alipour - Student Research Committee, Tabriz University of Medical Sciences, Tabriz, Iran

Background and aims: Osteoporosis is a silent disease which may not be detected until bonefractures appear. Approximately ۵۰% of women over the age of ۵۰ will suffer from a fragilityfracture in their remaining lifetime. Early detection is the only prevention of osteoporosis. Artificialintelligence (AI) has been used as a specified technology to interpret images in the diagnosisof many disorders, such as diabetic retinopathy. New studies reported that recent developments inAI have led to successful applications in the diagnosis of osteoporosis.Method: We conducted a comprehensive systematic search of Medline (via PubMed), Scopus,Embase, and Web of Science from inception to March ۲۰۲۳. The search was performed usingMeSH and free keywords such as “Artificial intelligence”, “Machine learning”, “Osteoporosis”and “DXA”. The database search also included gray literature and manual search. Two independentinvestigators screened located articles in multiple levels of title, abstract, and full-text. A thirdreviewer was involved in case of disagreements. We included studies that used AI models or machinelearning (ML) algorithms to diagnose osteoporosis (T-score < -۲.۵) using any type of medicalimaging with the gold standard (DXA) used as a reference/standard test. Two independentresearchers assessed the quality and bias of the studies that met our inclusion criteria accordingto the Quality Assessment of Diagnostic Accuracy Studies (QUADAS-۲) checklist. Microsoft OfficeExcel ۲۰۲۱ software is used to extract data such as ML algorithms, the regions and types ofimaging, and the accuracy of models. Pooled Accuracy was calculated using CMA v.۳.۷ software,using random effects model.Results: A systematic search of databases yielded ۵۹۰ articles. A total of ۲۹۷ articles were duplicates,and ۲۱۴ were excluded after reviewing the title and abstract of the articles. After reviewingthe full-text articles, ۵۳ articles were excluded. Finally, ۲۶ studies were included in this study. Inthe present review, most studies had a low bias, but depending on the type of studies, the outcomeReporting Bias was tangible in the studies. Radiographs and CT scans are the most common.Spine, hip, chest, and periapical images are the most popular among studies. About ۵۸% of studiesused SVM and CNN algorithms, and four of them used more than two ML algorithms. Overall,studies that used spine CT or radiographs, or dental radiographs had higher accuracy than others.A total of ۷۲,۹۵۷ images from ۲۶ studies were included in the meta-analysis. The pooled diagnosticaccuracy of osteoporosis was ۸۸.۶% (۹۵% CI: ۰.۰۱۹-۰.۰۲۱, P < ۰.۰۰۱).Conclusion: AI-based systems such as CNN and SVM have the potential to diagnose osteoporosisvia medical images such as radiographs and CT scans of some specific regions of the bodysuch as the spine and hip.However, there were some flaws in the development of AI-based real-world screening tools, suchas patient selection and methodological defects. Further studies by resolving the mentioned flawsare required to make AI a powerful and reliable tool for osteoporosis diagnosis.

Osteoporosis, Machine Learning (ML), Dual-energy X-ray absorptiometry (DXA), Diagnosis