Development and Evaluation of an Artificial Intelligence Model for Bone Mineral Density Prediction From X-Ray Images

This study aims to develop a robust artificial intelligence (AI) model for predicting Bone Mineral Density (BMD) from X-ray images using deep learning techniques, with a particular focus on improving the model's generalizability across diverse populations. The purpose is to provide an accessible, non-invasive screening tool for osteoporosis, reducing dependency on dual-energy X-ray absorptiometry (DEXA) scans, which are often unavailable or unaffordable in low-resource settings such as Bangladesh. Leveraging the convolutional neural network (CNN) architecture, this AI model is expected to assist in early osteoporosis diagnosis and management, ultimately improving clinical decision-making and healthcare efficiency.

This case-control observational study will be conducted in the Radiology Department of Ibn Sina Diagnostic and Consultation Center, Uttara. The study comprises both prospective and retrospective data collection phases, allowing for comprehensive data aggregation. During the prospective phase, data will be collected directly from eligible patients undergoing X-ray imaging and DEXA scans. For the retrospective phase, historical data will be extracted from clinical databases, including X-ray images and corresponding BMD reports. The study aims to address variations in bone health across a broad demographic, reflecting the prevalence of osteoporosis among different ages, genders, and clinical backgrounds in Bangladesh.

In Bangladesh, osteoporosis remains underdiagnosed due to the limited availability of DEXA scanners and trained personnel, particularly in rural and resource-constrained areas. The standard diagnostic pathway often begins with symptomatic X-ray imaging, followed by a DEXA scan if osteoporosis is suspected. This two-step process is costly and time-consuming, delaying diagnosis and treatment, which can lead to serious complications, including fractures. AI-driven predictions of BMD from X-ray images have the potential to streamline this pathway, enabling cost-effective screening and prioritization of patients who may need further DEXA-based testing. The AI model will be trained using a comprehensive dataset that includes demographic and clinical covariates-such as age, gender, menopausal status, and comorbid conditions like diabetes and cardiovascular disease-capturing correlations that could enhance prediction accuracy. Ultimately, the goal is to offer a reliable, scalable solution for osteoporosis screening that could be integrated into existing clinical workflows and alleviate the need for DEXA in settings where it is unavailable.

The study targets a diverse population group, including individuals with normal bone density, osteopenia, and osteoporosis as defined by DEXA measurements. This diversity ensures that the AI model can account for a wide spectrum of patient profiles and enhance its predictive robustness. The population will consist of adults across all age groups and genders, including both symptomatic and asymptomatic individuals.

The study will follow a structured protocol for data collection, aiming to gather comprehensive information on patients that may influence bone health. Key variables will include demographic details such as age, gender, and menopausal status; clinical variables like the presence of comorbidities such as diabetes and cardiovascular disease, BMI, and history of fractures; and imaging and diagnostic results, specifically X-ray images (spine or hip) and DEXA scan results for ground truth BMD values. In the prospective phase, eligible patients undergoing X-ray or DEXA scans will be approached for consent, and upon agreement, their clinical and demographic data will be recorded, including a unique identifier to ensure data integrity and confidentiality. Anonymized X-ray and DEXA images will then be collected, forming the primary dataset for AI training. The retrospective phase will involve data extraction from existing clinical records, focusing on spine and hip X-ray images and corresponding BMD results. Identifiable patient information will be removed to protect privacy. This historical dataset will complement the prospective data, providing a broader spectrum of cases and contributing to model generalizability.