Opportunistic screening for osteoporosis using chest X-rays and deep learning: A systematic review and meta-analysis.

PURPOSE: To evaluate the diagnostic performance of deep learning (DL) algorithms applied to chest radiographs (CXR) for detecting osteoporosis and assess their potential for clinical implementation.

METHODS

A systematic review and meta-analysis was conducted including studies that validated convolutional neural network (CNN)-based DL models to detect osteoporosis from CXRs. Exclusion criteria included studies using imaging other than CXR and non-DL models. Quantitative synthesis included pooled sensitivity, specificity, and construction of a summary receiver operating characteristic (SROC) curve. Model quality was evaluated using the APPRAISE-AI framework.

RESULTS

Out of 1401 records screened, 12 studies met inclusion criteria, comprising a total of 200,796 patients and 20 unique DL models. For internal validation of 13 DL models, the pooled sensitivity was 0.85 [95% CI: 0.82-0.87] and specificity of 0.83 [95% CI: 0.81-0.85]. The SROC analysis of all 20 DL models yielded a summary AUC of 0.69 [95% CI: 0.68-0.71]. A total of six studies performed external validation, with AUCs ranging from 73.4% to 94.6%, and specificity ranging from 74.2% to 90.9%, aside from one outlier. Heterogeneity was high (I2 > 98%) and driven primarily by country of origin. The APPRAISE-AI assessment indicated nine studies were high quality, supporting their use as a clinical decision support tool, while the others were of moderate quality.

CONCLUSION

DL models applied to CXR show promising diagnostic performance for opportunistic osteoporosis screening. However, substantial heterogeneity in internal validation and the limited number of externally validated studies underscore the need for further research to improve generalizability and support real-world clinical implementation.