Semi-supervised Learning for Bone Mineral Density Estimation in Hip X-Ray Images

Bone mineral density (BMD) is a clinically critical indicator of osteoporosis, usually measured by dual-energy X-ray absorptiometry (DEXA). Due to the limited accessibility DEXA machines and examinations, osteoporosis often under-diagnosed under-treated, leading increased fragility fracture risks. Thus it highly desirable obtain BMDs with alternative cost-effective more accessible medical imaging examinations such as plain films. In this work, we formulate BMD estimation from hip images regression problem. Specifically, propose new semi-supervised self-training algorithm train model using coupled unlabeled pseudo BMDs. Pseudo are generated refined iteratively for during self-training. We also present novel adaptive triplet loss improve model’s accuracy. On an in-house dataset 1,090 (819 unique patients), our method achieves high Pearson correlation coefficient 0.8805 ground-truth It offers good feasibility use cheaper opportunistic screening.