Researchers at Seoul National University Hospital (SNUH) have developed an AI model capable of both classifying bone density and explaining its reasoning using chest X-ray (CXR) images. The system can determine whether a patient has normal bone density, osteopenia, or osteoporosis by analyzing clinically relevant features—such as the spine and ribs—and comparing them with patterns learned during training.
The team used data from 14,502 women who received both CXR and DXA bone density tests at SNUH between 2004 and 2019. Four foundation models—two trained on general images and two on medical images—were tested using three validation methods: linear validation, partial fine-tuning, and low-rank adaptation. One general-image model, DINOv2, achieved the strongest performance when paired with low-rank adaptation, reaching an AUC of 93%, according to results published in Osteoporosis International.
To address the “black box” issue common in medical AI, researchers also built an explainability system that highlights the image regions the model relies on and verifies whether those areas correspond to clinically meaningful bone structures. The team emphasized that high accuracy alone is not enough for clinical adoption—trust, interpretability, and multidimensional evaluation are essential.
The study suggests that AI could enable opportunistic osteoporosis screening using standard chest X-rays, helping detect low bone mass early without requiring a separate DXA scan. SNU researchers say their work provides criteria and guidance for evaluating and choosing foundation models for medical use.
Globally, interest in osteoporosis-detection AI is growing. Korea’s Promedius has regulatory approval for an AI tool that identifies osteoporosis from CXRs, while Taiwan’s Acer Medical has approval in Indonesia for similar technology. Singapore’s National University Health System has also developed an AI system that flags hypercalcemia, a condition linked to osteoporosis.
Click here to read the original news story.