IMPACT OF PROJECTION VIEWS (PA VS. AP) ON AI-BASED OSTEOPOROSIS SCREENING USING CHEST RADIOGRAPHS

G. Lee1, J. Park1, K. W. Kim 1, Y.-C. Ha2

1Promedius Inc., Seoul, South Korea,2Department of Orthopaedic Surgery, Seoul Bumin Hospital, Seoul, South Korea

Objective: To evaluate the diagnostic performance of an AI-based osteoporosis screening

model using chest radiographs (CXRs), with a specific focus on performance differences

between posterior–anterior (PA) and anterior–posterior (AP) views in a real-world clinical

setting.

Materials and Methods: This retrospective study included CXRs from 969 patients acquired

at a specialized musculoskeletal hospital in Seoul between July 2021 and December 2024. The

dataset comprised 555 PA images (osteoporosis prevalence 31.35%, female 68.83%, age

69.55 ± 10.04 years) and 414 AP images (prevalence 50.00%, female 71.01%, age 73.87 ±

12.17 years). All CXRs were paired with dual-energy X-ray absorptiometry (DXA) results

obtained within a 3-month interval. An AI model originally trained using PA CXRs was

evaluated without retraining. Diagnostic performance was assessed using AUROC, sensitivity,

specificity, accuracy, and F1-score.

Results: The model demonstrated higher diagnostic performance on PA images than on AP

images. For PA views, the AUROC was 0.8618, with a sensitivity of 91.95% and a specificity of

56.69%. For AP views, the AUROC was 0.7480, with a sensitivity of 56.04% and a specificity of

78.74%. Overall accuracy was comparable between PA and AP views (67.75% vs. 67.39%), and

the F1-score showed only a modest difference (PA: 0.6413; AP: 0.6321).

Conclusion: An AI-based osteoporosis screening model trained on PA chest radiographs

demonstrated robust performance in standard PA imaging and preserved clinically meaningful

diagnostic capability when externally validated on AP views. These findings support the

feasibility of opportunistic osteoporosis screening using CXRs in heterogeneous real-world

settings. Incorporating AP-specific training or calibration strategies may further enhance

performance and facilitate broader clinical deployment across diverse acquisition protocols.