Deep-learning framework for osteoporosis screening on low-dose X-rays: Addressing image quality variability and cross-ethnic database Heterogeneity.

The AIXA Osteo model architecture (X1AI-Osteo) has undergone internal validation in Taiwan, demonstrating its capability to evaluate osteoporosis and predict T-scores reliably. Nonetheless, its efficacy in alternative clinical environments has not been confirmed yet. This study uses real-world data from a Japanese clinical environment to externally evaluate the AIXA Osteo model and improves upon common image capture discrepancies in cross-domain applications and diagnostic variability arising from disparate reference databases. We used a structure-preserving CycleGAN and a large-mask inpainting model to enhance the quality of radiographs and restored regions with missing areas. Additionally, the T-scores between NHANES III and Asian standards were reconciled through a feature fusion module. In a validation set of 300 participants, preprocessing improved the area under the curve of the model from 92.9% to 97.2%, sensitivity from 88.6% to 95.7%, and the positive predictive value from 80.5% to 90.5%. With respect to the T-score prediction performance, the consistency correlation coefficient between model and dual-energy X-ray absorptiometry measurements improved from 0.956 to 0.994. These findings indicate that the proposed framework for image preprocessing and the feature fusion module support the application of X-ray-based artificial intelligence for osteoporosis screening in heterogeneous clinical environments.