Abstract

Purpose

To develop and validate an AI algorithm that enables dual-site (spine and hip) BMD assessment for opportunistic osteoporosis screening using a single kidney-ureter-bladder (KUB) radiograph.

Materials and methods

In this institutional review board approved prospective study, we developed the SHield pipeline for opportunistic osteoporosis screening. From an initial dataset of 15,175 KUB images, a final cohort of 4,436 patients (mean age 69.0 ± 12.2 years) was included for model development after exclusions for suboptimal quality or incomplete region of interests. The pipeline analyzes both the spine and hip regions on a single KUB to predict site-specific BMD values. Finally, it integrates these dual-site predictions to determine the lowest T-score, mimicking the standard clinical DXA reporting protocol which evaluates both the lumbar spine and proximal femur.

Results

On the internal test set (628 patients), the hip and spine AI models demonstrated Pearson correlation coefficients of 0.887 and 0.921 and RMSEs of 0.057 and 0.062, respectively, compared to DXA-measured BMD, achieving AUCs of 0.894 and 0.956 for predicting T-scores ≤-2.5. In a subsequent prospective feasibility study (51 patients), we employed a conservative detection threshold of Tm-score ≤-2.8 to minimize false positives and unnecessary referrals. The final AI pipeline achieved a Pearson correlation of 0.959, an AUC of 0.969, 100% PPV, 100% specificity, and 83.8% NPV.

Conclusion

This paper presents the first AI pipeline that enables gold-standard-aligned, dual-site BMD estimation and opportunistic osteoporosis screening from a single KUB radiograph. Validated prospectively, its high accuracy and specificity confirm its potential to significantly increase early detection rates without overwhelming clinical workflows.