An AI-based pipeline for osteoporosis/osteopenia prediction using hip radiographs

To develop and validate an artificial intelligence-based tool for the diagnosis of osteoporosis/osteopenia using hip radiographs. The tool aims to classify femurs into risk-based categories for osteoporosis/osteopenia, enabling patient prioritization, enhancing preventive medicine through incidental detection, and assisting clinicians' diagnosis in general.

The AI tool was designed to perform three preprocessing tasks before the osteoporosis/osteopenia prediction: (1) splitting images into single femurs, (2) identifying and discarding femurs with prostheses, and (3) cropping images to isolate the proximal femur. A total of 2691 anteroposterior hip radiographs from 1654 patients were included in the study. The osteoporosis/osteopenia prediction model was trained on 3227 single femur images and tested on 826. Additionally, a final evaluation experiment was conducted on 313 new radiographs from 239 patients to assess the tool's applicability.

The tool demonstrated high performance in the preprocessing tasks, achieving 99.0% accuracy in classifying single vs. double femur images, 99.3% accuracy in identifying prosthetic femurs, and 99.2% pixel accuracy in delineating the proximal femur before cropping. The final prediction model achieved an area under the curve of 86.6% for detecting osteoporosis/osteopenia in the test set and 81.0% in the final evaluation experiment.

The obtained results demonstrate the potential of the proposed AI-based pipeline for prediction of osteoporosis/osteopenia using hip radiographs. This study suggests that a tool based on the proposed methods could support DXA triage, incidental osteoporosis detection, and clinical decision-making in settings with limited access to bone densitometry.