Summary We evaluated foundation models for opportunistic osteoporosis screening from chest X-rays using a novel explainability framework. DINOv2 with low-rank adaptation achieved the best performance (AUC 0.93) while demonstrating clear clinical reasoning. Our findings highlight that explainability should be prioritized alongside accuracy in medical AI, enhancing trust in clinical deployment. Purpose Deep learning models show promise for opportunistic osteoporosis screening from chest X-rays but have traditionally relied on convolutional neural networks with limited explainability. This study introduces a quantitative framework for explainability evaluation and systematically compares diverse foundation models to identify an optimal balance between performance and explainability. Methods In this retrospective study, a retrospective dataset comprising 21,031 chest X-rays paired with bone mineral density scores from 14,502 female patients at Seoul National University Hospital was used. Twelve foundation model variants—combinations of natural- and medical-domain models fine-tuned using various strategies—were trained to classify osteoporosis status (normal, osteopenia, or osteoporosis). Foundation models were evaluated based on predictive performance (AUC, accuracy, sensitivity, and specificity) and explainability, assessed through occlusion analysis (AUC change after bone perturbation, <InlineEquation ID="IEq1"> <EquationSource Format="TEX">$\Delta _{bone}$</EquationSource> <EquationSource Format="MATHML"><math> <msub> <mi mathvariant="normal">Δ</mi> <mrow> <mi mathvariant="italic">bone</mi> </mrow> </msub> </math></EquationSource> </InlineEquation>) and saliency-map analysis (overlap between bone regions and saliency maps, <InlineEquation ID="IEq2"> <EquationSource Format="TEX">$\text {IoU}_{bone}$</EquationSource> <EquationSource Format="MATHML"><math> <msub> <mtext>IoU</mtext> <mrow> <mi mathvariant="italic">bone</mi> </mrow> </msub> </math></EquationSource> </InlineEquation>). Results DINOv2, fine-tuned with low-rank adaptation, achieved the highest predictive performance (AUC of 0.93; 95% CI, 0.92–0.94) and demonstrated robust explainability by focusing on clinically relevant bone structures, such as the spine and ribs. In osteoporosis screening from chest X-rays, statistical analysis showed that medical foundation models did not consistently outperform natural-domain models, and higher performance did not always correlate with better explainability. Conclusion Our findings underscore the necessity of incorporating explainability as a key criterion when selecting deep learning models for opportunistic osteoporosis screening. Furthermore, the proposed framework can be readily extended to other medical tasks, fostering the development of more trustworthy and interpretable AI-assisted screening tools.

Explainable opportunistic osteoporosis screening from chest X-rays: a retrospective comparison of foundation models

期刊论文
出版模式：

开放获取
发表日期：

2025年10月27日

查看全文

Jaewon Kim,
Sangmin Kwak,
Hyeokjong Lee,
Jooyoung Chang,
Sang Min Park

摘要

Summary

We evaluated foundation models for opportunistic osteoporosis screening from chest X-rays using a novel explainability framework. DINOv2 with low-rank adaptation achieved the best performance (AUC 0.93) while demonstrating clear clinical reasoning. Our findings highlight that explainability should be prioritized alongside accuracy in medical AI, enhancing trust in clinical deployment.

Purpose

Deep learning models show promise for opportunistic osteoporosis screening from chest X-rays but have traditionally relied on convolutional neural networks with limited explainability. This study introduces a quantitative framework for explainability evaluation and systematically compares diverse foundation models to identify an optimal balance between performance and explainability.

Methods

In this retrospective study, a retrospective dataset comprising 21,031 chest X-rays paired with bone mineral density scores from 14,502 female patients at Seoul National University Hospital was used. Twelve foundation model variants—combinations of natural- and medical-domain models fine-tuned using various strategies—were trained to classify osteoporosis status (normal, osteopenia, or osteoporosis). Foundation models were evaluated based on predictive performance (AUC, accuracy, sensitivity, and specificity) and explainability, assessed through occlusion analysis (AUC change after bone perturbation, $\Delta _{bone}$ $Δ_{bone}$ ) and saliency-map analysis (overlap between bone regions and saliency maps, $\text {IoU}_{bone}$ ${IoU}_{bone}$ ).

Results

DINOv2, fine-tuned with low-rank adaptation, achieved the highest predictive performance (AUC of 0.93; 95% CI, 0.92–0.94) and demonstrated robust explainability by focusing on clinically relevant bone structures, such as the spine and ribs. In osteoporosis screening from chest X-rays, statistical analysis showed that medical foundation models did not consistently outperform natural-domain models, and higher performance did not always correlate with better explainability.

Conclusion

Our findings underscore the necessity of incorporating explainability as a key criterion when selecting deep learning models for opportunistic osteoporosis screening. Furthermore, the proposed framework can be readily extended to other medical tasks, fostering the development of more trustworthy and interpretable AI-assisted screening tools.