Generative Artificial Intelligence for Opportunistic Osteoporosis Screening Using Lumbar Spine Imaging: Diagnostic Performance, Reproducibility, Effects of Few-Shot Prompting, and Comparison with Physicians

Abstract Purpose Osteoporosis remains underdiagnosed despite widespread availability of lumbar radiographs and computed tomography (CT). We compared physicians and multimodal generative artificial intelligence (AI) for opportunistic osteoporosis screening, focusing on diagnostic accuracy, reproducibility, and effects of few-shot prompting. Methods Forty-five lateral lumbar radiographs and 45 sagittal CT images were classified using DXA T-score thresholds of −1.0 and −2.5 as reference standards. Performance was compared among 42 board-certified orthopedic surgeons, 50 non-specialists, and ChatGPT, Gemini 3.1 Pro, and Claude 4 Opus. Each AI model evaluated all images five times under zero-shot and few-shot conditions, with 20 labeled reference cases provided for few-shot prompting. An exploratory analysis evaluated combined specialist-AI classifications using OR and AND rules. Results At the −1.0 cutoff, specialists outperformed non-specialists on radiographs (P<0.001). Claude achieved the highest radiographic accuracy (73.3%), followed by Gemini (64.4%), specialists (57.4%), ChatGPT (55.6%), and non-specialists (44.0%). Educational examples did not improve physician accuracy, and few-shot prompting did not consistently improve AI accuracy. However, reproducibility improved in several models, most notably Claude. An OR rule combining specialist and Claude classifications increased sensitivity to 80.0% on radiographs and 90.0% on CT, with reduced specificity. Conclusion Generative AI achieved physician-comparable performance under selected conditions but showed model-dependent variability. Few-shot prompting may improve consistency without improving accuracy. Exploratory combined assessment suggested complementary physician-AI errors may reduce missed cases at the cost of false-positive classifications. Generative AI may therefore serve as a decision-support tool for opportunistic osteoporosis screening but should complement rather than replace physician judgment.