# Lumbar Bone Mineral Density Estimation from Chest X-ray Images: Anatomy-aware Attentive Multi-ROI Modeling

**Authors:** Fakai Wang, Kang Zheng, Le Lu, Jing Xiao, Min Wu, Chang-Fu Kuo, Shun Miao  
**arXiv:** [2201.01838](https://arxiv.org/abs/2201.01838)  
**Date:** 2022-01-05  
**Summary page:** https://www.summarizepaper.com/en/paper/2201.01838v2/

## Key points

- Proposed method: Predicting bone mineral density (BMD) from chest X-ray (CXR) images
- CXR images are more accessible and cost-effective compared to traditional BMD examinations
- Utilizes deep learning models, specifically the anatomy-aware Attentive Multi-ROI model
- Extracts local bone textures and generates robust feature representations
- Automatically detects Regions of Interest (ROIs) of local CXR bone structures
- Employs a transformer-based encoder to exploit both local and global information for accurate BMD estimation
- Evaluated on a dataset of 13,719 CXR patient cases with ground truth BMD measured by Dual-energy X-ray Absorptiometry (DXA)
- Strong correlation between predicted BMD and ground truth, with a Pearson correlation coefficient of 0.894 on lumbar 1
- High classification performance in osteoporosis screening, with an average Area Under the Curve (AUC) of 0.968
- Potential for early osteoporosis screening and public health promotion
- Overcomes limitations in access to DXA scans by leveraging commonly accessible medical imaging examinations like CXR
- Highlights potential utility of plain film or X-ray as an alternative screening tool for osteoporosis alongside existing techniques like DXA and quantitative computed tomography (QCT)
- Provides valuable insights into using deep learning models for estimating lumbar vertebra BMD from chest X-ray images
- Promising clinical potential for opportunistic screening and implications for enhancing public health initiatives related to osteoporosis detection and prevention

## Summary

In this paper, the authors propose a method to predict bone mineral density (BMD) from chest X-ray (CXR) images which are more accessible and cost-effective compared to traditional BMD examinations. The method utilizes deep learning models, specifically the anatomy-aware Attentive Multi-ROI model, to extract local bone textures and generate robust feature representations. The model first automatically detects Regions of Interest (ROIs) of local CXR bone structures and then employs a transformer-based encoder to exploit both local and global information in the image for accurate BMD estimation. The proposed method is evaluated on a dataset of 13,719 CXR patient cases with ground truth BMD measured by Dual-energy X-ray Absorptiometry (DXA), which is considered the gold standard. The results show that the predicted BMD has a strong correlation with the ground truth, with a Pearson correlation coefficient of 0.894 on lumbar 1. Furthermore, when applied in osteoporosis screening, the model achieves high classification performance with an average Area Under the Curve (AUC) of 0.968. This study represents the first effort to use CXR scans for predicting BMD and holds great potential for early osteoporosis screening and public health promotion. By leveraging commonly accessible medical imaging examinations like CXR, this algorithm can overcome limitations in access to DXA scans and contribute to improving diagnosis and treatment of osteoporosis. Additionally, it highlights the potential utility of plain film or X-ray as an alternative screening tool for osteoporosis alongside existing techniques like DXA and quantitative computed tomography (QCT). Overall, this research provides valuable insights into using deep learning models for estimating lumbar vertebra BMD from chest X-ray images. The proposed algorithm demonstrates promising clinical potential for opportunistic screening and has implications for enhancing public health initiatives related to osteoporosis detection and prevention.