Development and validation of a machine learning-derived radiomics model for diagnosis of osteoporosis and osteopenia using quantitative computed tomography

Qianrong Xie, Yue Chen, Yimei Hu, Fanwei Zeng, Pingxi Wang, Lin Xu, Jianhong Wu, Jie Li, Jing Zhu, Ming Xiang, Fanxin ZengBMC Medical Imaging 2022 · DOI

TL;DR

This study introduces a novel machine learning model that combines quantitative computed tomography (QCT) radiomics features with clinical data to diagnose osteoporosis and osteopenia. The developed model demonstrates superior diagnostic performance compared to clinical data alone, offering a more accurate and potentially earlier detection tool for these bone health conditions.

What Problem Does This Paper Address?

Osteoporosis and osteopenia are significant musculoskeletal diseases characterized by reduced bone mass and increased fracture risk, posing a growing global health challenge, particularly with aging populations. While dual-energy X-ray (DXA) is a common diagnostic tool, it can be affected by factors like vascular calcification and body position. Quantitative computed tomography (QCT) offers a more sensitive measure of bone mineral density (BMD) and can assess bone microstructure. However, traditional QCT analysis often relies on limited parameters. This paper addresses the need for more comprehensive and accurate diagnostic tools by leveraging the power of radiomics – extracting a large number of quantitative features from medical images – to improve the discrimination of osteoporosis and osteopenia.

Methods

The researchers employed a multi-step machine learning approach. First, they extracted a comprehensive set of 851 radiomics features from QCT images of the third lumbar vertebra (L3). To manage this high dimensionality and select the most informative features, they utilized a two-stage process:

  1. Minimum Redundancy Maximum Relevance (mRMR): This technique pre-selected 20 potential features by prioritizing those with high relevance to the diagnosis while minimizing redundancy among them.
  2. Least Absolute Shrinkage and Selection Operator (LASSO): LASSO was then applied to further reduce the feature set, selecting 6 optimal radiomics features with non-zero coefficients that effectively discriminated between osteoporosis and osteopenia.

A combined clinical-radiomic model was then constructed by integrating these selected radiomics features with key clinical characteristics (age, alkaline phosphatase, and homocysteine). The performance of this combined model, as well as separate clinical and radiomics models, was evaluated using receiver operating characteristic (ROC) curves, accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV).

Datasets

The study retrospectively collected QCT data from 635 patients at Dazhou Central Hospital between November 2016 and November 2019. These patients were classified into normal, osteopenia, and osteoporosis groups based on their T-scores. The dataset was divided into a training cohort (N=414) and a test cohort (N=176) in a 7:3 ratio for model development and validation. No new benchmark datasets were introduced; the study utilized existing patient data.

Key Insights and Ideas

  • Radiomics as a powerful supplement to clinical data: The study strongly suggests that radiomics features extracted from QCT images can significantly enhance the diagnostic capabilities beyond traditional clinical markers.
  • L3 vertebra as a key diagnostic site: The selection of the third lumbar vertebra (L3) for radiomics analysis proved effective, aligning with previous research highlighting its specificity for bone loss assessment due to its high metabolic rate and lower bone density.
  • The synergy of multi-dimensional data: The combined clinical-radiomic model outperformed models based on either clinical data or radiomics alone, underscoring the benefit of integrating diverse data sources for improved diagnostic accuracy.
  • Identification of key radiomics features: The study successfully identified specific radiomics features (including wavelet-based and first-order features) that are highly discriminative for osteoporosis and osteopenia, providing insights into the underlying bone microstructure changes.

Key Results

The combined clinical-radiomic model achieved an impressive Area Under the Curve (AUC) of 0.96 (95% CI, 0.95–0.98) in the training cohort and 0.96 (95% CI, 0.92–1.00) in the test cohort. This performance was significantly superior to the clinical model alone, which yielded AUCs of 0.81 (training) and 0.79 (test). The radiomics model itself also performed very well, with AUCs of 0.96 (training) and 0.96 (test), indicating it was not inferior to the combined model. The nomogram developed from the combined model demonstrated good agreement between predicted and observed probabilities, as confirmed by calibration curves. Decision curve analysis also indicated that the combined model offered greater net benefit compared to the clinical model across a range of threshold probabilities.

Limitations and Future Work

The authors acknowledge several limitations:

  • Single-center study: The data was collected from a single institution, which may lead to an overly optimistic evaluation of the model's performance. External validation with data from multiple centers is recommended.
  • Lack of follow-up data: The study did not incorporate follow-up data such as therapeutic response or fracture rates, which could further validate the clinical utility of the model.

Future work could focus on external validation to confirm the generalizability of the model and explore its application in predicting fracture risk or monitoring treatment response. The study also suggests that this radiomics-derived rad-score could serve as an independent risk factor for patients with abnormal BMD, potentially guiding clinical decision-making.

Summary updated

BibTeX
@article{xie2022development,
  author = {Qianrong Xie and Yue Chen and Yimei Hu and Fanwei Zeng and Pingxi Wang and Lin Xu and Jianhong Wu and Jie Li and Jing Zhu and Ming Xiang and Fanxin Zeng},
  title = {{Development and validation of a machine learning-derived radiomics model for diagnosis of osteoporosis and osteopenia using quantitative computed tomography}},
  journal = {BMC Medical Imaging},
  year = {2022},
  doi = {10.1186/s12880-022-00868-5},
}
Figure