Explainable Multi-Class Osteoporosis Detection and Classification from Knee X-Ray Images Using an ROI-Guided Dual-Branch Deep Learning Framework

This paper presented ORSA-Net, an explainable ROI-guided dual-branch deep learning framework for detecting and classifying osteoporosis from knee X-ray images. The framework classifies radiographs into normal, osteopenia, and osteoporosis categories and combines bone segmentation, self-supervised learning, CNN-based local feature extraction, transformer-based global representation learning, ordinal clas-sification, optional clinical-feature fusion, and uncertainty es-timation. The proposed method addresses several limitations iden-tified in previous studies, including binary-only prediction, limited anatomical localization, class imbalance, absence of severity-aware learning, weak confidence calibration, and in-sufficient external validation. Grad-CAM explanations and uncertainty-based referral logic are included to improve trans-parency and clinical safety. The system is intended for low-cost opportunistic screening and clinical decision support, particularly in settings where DXA equipment is not readily available. However, it should not replace DXA measurements, radiologist interpretation, or specialist assessment. Actual clinical usefulness must be estab-lished through patient-independent testing, external validation, prospective studies, and ethical approval.

[1]     C. S. Ho et al., “Application of deep learning neural network in predicting bone mineral density from plain X-ray radiography,” Archives of Osteoporosis, vol. 16, 2021, doi: 10.1007/s11657-021-00985-8. [2]     Y. Sato et al., “Deep learning for bone mineral density and T-score prediction from chest X-rays: A multicenter study,” 2022. Available through PubMed, PMID: 36140424. [3]     K. Golestan, C. A. Syme, A. Bilbily, S. Zuberi, M. Volkovs, T. Poutanen, and M. D. Cicero, “Approximating femoral neck bone mineral density from hand, knee, and pelvis X-rays using deep learning,” Journal of Medical Artificial Intelligence, vol. 6, 2023, doi: 10.21037/jmai-23-10. [4]     A. Boonrod et al., “Deep learning for osteoporosis screening using anteroposterior hip radiographs,” 2024. Available through PubMed, PMID: 38896146. [5]     T. Asamoto et al., “External validation of a deep learning model for predicting bone mineral density on chest radiographs,” 2024. Available through PubMed, PMID: 38472499. [6]     U. Hwang, C.-H. Lee, and K. Yoon, “Osteoporosis prediction from hand X-ray images using segmentation-for-classification and self-supervised learning,” arXiv preprint arXiv:2412.05345, 2024. [7]     Y. Yoshida, Y. Sato, Y. Kai, Y. Hirono, T. Sato, and T. Kasai, “Utility of osteoporosis screening based on estimation of bone mineral density using bidirectional chest radiographs with deep learning models,” 2025. Available through PubMed, PMID: 40206487. [8]     J. Tang, X. Yin, J. Lai, K. Luo, and D. Wu, “Fusion of X-ray images and clinical data for a multimodal deep learning prediction model of osteoporosis: Algorithm development and validation study,” JMIR Medical Informatics, 2025, doi: 10.2196/70738. [9]     P. Rajpurkar et al., “MURA: Large dataset for abnormality detection in musculoskeletal radiographs,” arXiv preprint arXiv:1712.06957, 2017. [10]    I. M. Wani and S. Arora, “Knee X-ray osteoporosis database,” Mendeley Data, Version 2, 2021. [Online]. Available: https://data.mendeley.com/ datasets/fxjm8fb6mw/2.